논문 다이제스트Paper Digest

Artificial intelligence based methods for anatomy segmentation can help automate characterization of large imaging datasets. The growing number of similar functionality models raises the challenge of evaluating them on datasets that do not contain ground truth annotations. We introduce a practical framework to assist in this task. We harmonize the segmentation results into a standard, interoperable representation, which enables consistent, terminology-based labeling of the structures. We extend 3D Slicer to streamline loading and comparison of these harmonized segmentations and demonstrate how standard representation simplifies review of the results using interactive summary plots and browser-based visualization using the OHIF Viewer. To demonstrate the utility of the approach, we apply it to evaluating segmentation of 31 anatomical structures (lungs, vertebrae, ribs, and heart) by 6 open-source models-TotalSegmentator 1.5 and 2.6, Auto3DSeg, MOOSE, MultiTalent, and CADS-for a sample of computed tomography scans from the publicly available National Lung Screening Trial dataset. We demonstrate the utility of the framework in enabling automating loading, structure-wise inspection, and comparison across models. Preliminary results ascertain the practical utility of the approach in allowing quick detection and review of problematic results. The comparison shows excellent agreement segmenting some (e.g., lung) but not all structures (e.g., some models produce invalid vertebrae or rib segmentations). The open-source resources developed include segmentation harmonization scripts, interactive summary plots, and visualization tools. These resources assist in segmentation model evaluation in the absence of ground truth, ultimately enabling informed model selection.

Depressive symptoms frequently co-occur in individuals with Mild Cognitive Impairment (MCI) and are thought to accelerate neurodegenerative progression. However, the underlying neural mechanisms of Depressed MCI (DMCI) remain largely unclear. This study employed a multimodal resting-state functional magnetic resonance imaging (rs-fMRI) approach combined with advanced machine learning techniques, to systematically examine spontaneous brain activity patterns and topological organization differences among DMCI, non-depressed MCI (nDMCI), and non-depressed cognitively normal controls (nDCN). The research analyzed amplitude-based rs-fMRI measures and graph-theoretical features. Voxel-wise analyses and connectivity comparisons were conducted between groups. Additionally, classification tasks were performed using classical machine learning models and a graph reinforcement learning (GRL) model. DMCI individuals exhibited increased activity in the right insula and decreased amplitude of low-frequency fluctuation (ALFF) in the left calcarine cortex, along with heightened fractional ALFF (fALFF) and percent amplitude of fluctuation (PerAF) in the precuneus and parahippocampal regions. Graph metrics revealed disrupted global and local efficiency in nDMCI compared to nDCN. Using differential matrices, machine learning achieved optimal accuracies of 0.82 ± 0.15 (DMCI vs. nDMCI) and 0.84 ± 0.15 (DMCI vs. nDCN). Conversely, the GRL model for nDMCI vs. nDCN peaked at 0.66 ± 0.02 using full matrices, dropping to 0.60 ± 0.04 with filtering, indicating deep graph models require complete topological data for subtle differences. Rs-fMRI and graph learning approaches offer promising avenues for subtype classification, highlighting the hyperactivity of the right insula and the integrity of the whole-brain functional connectivity matrix as crucial potential biomarkers of early pathological changes.

Tire and Road Wear Particles (TRWP) are pervasive environmental contaminants, yet the molecular mechanisms linking their toxic derivative, 6PPD-quinone (6PPD-Q), to bladder cancer (BLCA) progression remain obscure. This study integrates network toxicology with experimental validation to elucidate this complex pathogenicity. We screened six representative TRWP compounds and utilized a comprehensive machine learning framework involving 113 model combinations, identifying the Gradient Boosting Machine (GBM) as the optimal classifier. Crucially, SHAP interpretability analysis revealed RPS6KB2 (S6K2) as a pivotal risk driver, while molecular docking demonstrated that 6PPD-Q exhibits superior binding affinity (Binding energy = -7.405 kcal/mol) to S6K2 compared to its parent compound. In vitro assays confirmed that S6K2 is upregulated in BLCA and essential for malignancy. Exposure of BLCA cells to 6PPD-Q dose-dependently upregulated S6K2, significantly (p < 0.05) promoting proliferation, migration, and invasion as evidenced by EdU and Transwell assays. Notably, S6K2 silencing effectively reversed these 6PPD-Q-induced malignant phenotypes. These findings provide the first evidence that 6PPD-Q drives BLCA progression via the specific upregulation of S6K2, offering a novel theoretical basis for assessing the health risks of TRWP exposure.

Background: Colorectal cancer (CRC) remains the third most prevalent malignancy globally, with approximately 154,000 new cases and 54,000 projected deaths anticipated for 2025. The recent advancement of foundation models in computational pathology has been largely propelled by task-agnostic methodologies that overlook organ-specific crucial morphological patterns that represent distinct biological processes that fundamentally influence tumor behavior, therapeutic response, and outcomes. Methods: In this study, we develop a novel, interpretable AI model, PRISM (Prognostic Representation of Integrated Spatial Morphology), that incorporates a continuous variability spectrum within each distinct morphology to reflect the principle that malignant transformation occurs through incremental evolutionary processes. PRISM is trained on 15 million histological images extracted from surgical resection specimens of 2957 patients. Results: PRISM achieved superior prognostic performance for five-year OS (AUC = 0.70 ± 0.04; accuracy = 68.37% ± 4.75%; HR = 3.21, 95% CI = 2.18-4.72; p < 0.0001 using multi-variate cox-proportional hazards model), outperforming existing CRC-specific methods by 15% and AI foundation models by ~23% accuracy. It showed sex-agnostic robustness (AUC Δ = 0.02; accuracy Δ = 0.15%) and stable performance across clinicopathological subgroups, with minimal accuracy fluctuation (Δ = 1.44%) between 5FU/LV and CPT-11/5FU/LV regimens, replicating the Alliance cohort finding of no survival difference between treatments. Conclusions: These results establish PRISM as a promising, interpretable tool for AI-driven prognostication, with potential for future extension to other cancer types and stages.

Artificial intelligence (AI) and deep learning (DL) are transforming cancer research and clinical care, with histopathology playing a central role in this transformation. In colorectal cancer (CRC), the second leading cause of cancer mortality world-wide, multimodal and vision-language models (VLMs) hold particular promise for enhancing the standardisation of histopathology reporting, the understanding of disease biology, and the discovery of novel prognostic indicators. Despite the availability of guidelines and reporting templates for essential prognostic indicators, variability remains in how key features such as TNM staging or tumour deposits are assessed and reported in routine clinical practice. AI-based tools have the potential to support refined extraction of established and extended features directly from whole-slide images. In parallel, recent studies have shown that DL models applied to pathology slides and associated AI-based biomarkers can outperform traditional histopathological prognostic indicators and uncover novel parameters, including tumour-adipocyte interactions, tumour-stroma ratio, and immune cell patterns at the invasive margin. Here, we review recent advances in both domains: AI-assisted standardisation of CRC pathology reporting and AI-driven identification of novel prognostic biomarkers. We highlight the need to refine and standardise CRC reporting practices and propose that a harmonised approach combining established pathology features with AI-derived prognostic indicators could refine risk assessment and improve outcomes for CRC patients. © 2026 The Author(s). The Journal of Pathology published by John Wiley & Sons Ltd on behalf of The Pathological Society of Great Britain and Ireland.

In dyssynchronous heart failure (DHF), left bundle branch block (LBBB) causes inhomogeneous left ventricular (LV) workload and systolic dysfunction. We aimed to investigate underlying metabolic remodelling in an ovine model. Eleven sheep with dual-chamber-pacemakers for LBBB-like activation (DHF) were studied at baseline and after eight weeks. Six untreated sheep served as controls (CTRL). Regional workload was evaluated using invasive hemodynamics and echocardiography. 18F-fluorodeoxyglucose-tracer positron-emission-tomography/computed tomography visualized regional glucose-uptake. Magnetic resonance imaging assessed fibrosis (late gadolinium enhancement, LGE). Septal and lateral wall tissue was analysed with histology, confocal microscopy, ultra-high-performance liquid chromatography-high resolution mass-spectrometry (UHPLC-HRMS). Dyssynchrony induced low septal and high lateral asymmetry in workload and glucose-uptake. After eight weeks, DHF animals exhibited LV dilation and LVEF decline (31.1±5.1% vs. 59.4±3.5% at baseline, p<0.05). Septal thinning and lateral hypertrophy rebalanced workload and glucose-uptake. No fibrosis was seen on LGE or histology. DHF-animals showed enrichment of mitochondria at the intercalated discs (EMID-sign) - highest in the lateral wall (DHF septal 7.0±4.8% vs. lateral 48.4±12.3%, p<0.05). Mitochondrial redox balance in DHF shifted towards a more oxidised state without evidence of oxidative stress. Metabolomics revealed no differences between septal and lateral walls but severe energy depletion of tricarboxylic acid cycle substrates and phosphocreatine in DHF (fold change DHF/CTRL 0.01, p<0.01). Experimental DHF is characterised by non-fibrotic, dilated LV without signs of oxidative stress. Workload increase in the lateral wall leads to hypertrophy and EMID, homogenizing metabolic profiles between wall segments. However, the ventricle enters energy starvation and systolic dysfunction.

Peripheral neuropathies contribute to patient disability but may be diagnosed late or missed altogether due to late referral, limitation of current diagnostic methods and lack of specialized testing facilities. To address this clinical gap, we developed NeuropathAI, an interpretable deep learning-based multiclass classification system for rapid, automated diagnosis and differentiation of 88 patients with diabetic peripheral neuropathy (DPN), chemotherapy-induced peripheral neuropathy (CIPN), chronic inflammatory demyelinating polyneuropathy (CIDP), and human immunodeficiency virus-associated sensory neuropathy (HIV-SN). A deep learning-based multiclass system was developed to analyze corneal nerve images. These images were preprocessed to train and validate the proposed model and the diagnostic utility was evaluated from the accuracy, F1-score and area under the curve to derive sensitivity, specificity and precision. NeuropathAI achieved excellent results: AUC-96.75%, sensitivity-83.87%, specificity-95.07%, and demonstrated excellent discrimination for CIDP, CIPN, HIV-SN and DPN with one-vs-all AUC scores of 97%, 93.1%, 99.7% and 96.9%, respectively. Explainability visualization through heatmaps demonstrated that regions of decision making by the model localized to areas with nerve fiber loss, enhancing interpretability. NeuropathAI achieved rapid and accurate diagnosis of four of the most prevalent peripheral neuropathies globally, underscoring the potential of artificial intelligence-driven corneal image analysis for the rapid diagnosis and differentiation of peripheral neuropathies.

Research has demonstrated a broad network of dysfunction across the brain in attention-deficit/hyperactivity disorder (ADHD), suggesting the potential role of white matter (WM) organization. In this study, we sought to estimate the developmental trajectories of brain WM myelination in children with ADHD. Neuroimaging and clinical data were collected as part of a longitudinal community-based pediatric cohort (Nscans = 400; 195 with ADHD; age range = 9-14 years). Brain WM myelin was examined for 71 WM tracts across 3 time points using the T1-weighted (T1w)/T2-weighted (T2w) ratio. Tracts were defined via a deep learning-based automated tractography method, performed on participant diffusion-weighted images. Linear and nonlinear regression analyses were conducted to examine group differences in T1w/T2w ratio values. In addition to this, voxelwise analysis was undertaken at each time point. Brainwide, children with ADHD were found to exhibit the same developmental profile as children without ADHD for WM myelin. No group effects were seen at a cross-sectional or longitudinal level. Consistent with previous work, modeling suggests nonlinear developmental increases with age across most tracts. This nonlinear relationship was characterized by a positive parabolic or U-shaped developmental trajectory. These findings indicate that there may not be distinct differences in the development of brain WM myelination between children with and without ADHD. However, this suggests that previously reported differences in ADHD brain WM development may be attributable to properties other than myelin, such as fiber architecture and axon diameter. This further informs the understanding of brain development and highlights the need for more multimodal longitudinal work.

Robust automated classification of paramagnetic rim lesions (PRLs) and remyelinated lesions based on iron and myelin content in people with multiple sclerosis (pwMS). In this prospective study (2018-2022), three-dimensional (3D) brain magnetic resonance imaging (MRI) from 180 pwMS (mean age: 47 ± 14 years, 108 females) were acquired at 3T including fluid-attenuated inversion recovery, magnetization-prepared 2 rapid acquisition gradient echoes (MP2RAGE), and T2*-weighted segmented echo-planar imaging. Post-processed quantitative susceptibility mapping (QSM) and filtered phase unwrapped (PU) images were generated. Ground truth was established through expert manual rating. Performance was assessed using nested cross-validation (CV) including outer test sets. Three neural network configurations were evaluated: (1) PRL classification with QSM and MP2RAGE, (2) PRL classification with PU and MP2RAGE, and (3) multiple lesion phenotype (MLP) classification with QSM and MP2RAGE. For PRL classification, our network achieves a top mean validation F1 score of 0.737 ± 0.027 in the MP2RAGE-QSM configuration trained on PRLs, and a best test performance of 0.709 ± 0.040 when trained on MLP. The MP2RAGE-PU configuration yielded validation and test F1 scores of 0.733 ± 0.021 and 0.662 ± 0.037, respectively. The QSM-based MLP classification achieved a macro F1 test score of 0.728 ± 0.012. Deep learning can automate the classification of PRLs and QSM lesion phenotypes with high accuracy, potentially facilitating the detection of PRLs (included in the new diagnostic criteria) and remyelinated lesions, guiding personalized treatment decisions in pwMS.

Fast magnetic resonance imaging (MRI) can significantly improve patient tolerance and examination efficiency, but it may compromise image quality. This study investigated the feasibility of using a deep learning-assisted three-dimensional iterative image enhancement (DL-3DIIE) system to achieve high-resolution fast imaging of the knee. This prospective study included participants scheduled for knee MRI plain scans at the Northern Jiangsu People's Hospital between September 2023 and January 2024. The participants underwent knee MRI with both conventional and fast scans. Three MRI protocols were compared: conventional MRI, fast MRI with reduced acquisition parameters, and DL-3DIIE MRI, which enhanced fast images with deep learning to improve overall image quality. Image quality was assessed subjectively (quality scores) and objectively using peak signal-to-noise ratio (PSNR), multi-scale structural similarity index (MS-SSIM), signal-to-noise ratio (SNR), and contrast-to-noise ratio (CNR). The analysis included 134 patients (mean age 55.1±9.5 years; 46.3% male). For both sagittal proton density weighted imaging-fast spin echo (PDWI-FSE) and T1 weighted imaging-fast spin echo (T1WI-FSE) sequences, DL-3DIIE MRI achieved significantly higher SNRs and CNRs across all tissues and tissue contrasts compared with both conventional and fast MRI (all P<0.05). Quantitative metrics of image quality were also improved, with PSNR and MS-SSIM significantly higher for DL-3DIIE MRI than fast MRI in all sequences (all P<0.001). Subjective image quality assessment demonstrated that DL-3DIIE MRI yielded significantly higher scores for lesion conspicuity, margin delineation, and overall diagnostic confidence compared with conventional MRI (all P<0.05). DL-3DIIE MRI provides superior quantitative and subjective image quality compared with both conventional and fast MRI, including higher SNR, CNR, PSNR, and MS-SSIM, as well as improved lesion and margin visibility. These findings support the potential of DL-3DIIE to accelerate knee MRI while preserving or improving diagnostic performance.

Minimizing acquisition time and motion-artifacts remains challenging in magnetic resonance imaging (MRI) with demands on high-resolution images for accurate diagnosis and treatment. Deep learning-based image restoration offers promising solution by generating high-resolution and artifact-free MR images from low-resolution or motion-corrupted data. To facilitate practical deployment in clinical workflows, this study presents a time-/GPU-efficient framework using 2D network (TS-RCAN) for pseudo-3D MRI super-resolution reconstruction (SRR) and motion-artifact reduction (MAR). Optimal down-sampling factors were identified to balance SRR accuracy and acquisition time. MAR training used a standardized method to induce controllable motion-artifacts of varying severity. Network performance was benchmarked against state-of-the-art 3D networks. Results showed the down-sampling factor [Formula: see text] for [Formula: see text] acceleration and [Formula: see text] for [Formula: see text] acceleration achieved optimal SRR performance. TS-RCAN outperformed most 3D networks by > 0.01/1.5 dB in SSIM/PSNR while reducing GPU load and inference time by up to 90%. For MAR, TS-RCAN exceeded UNet by up to 0.014/1.48 dB in SSIM/PSNR. Additionally, uncertainty estimation correlated with image quality metrics, enabling accuracy prediction without ground truth. TS-RCAN provides an efficient, accurate framework for SRR and MAR with practical relevance to clinical MRI, and offers a flexible basis for future extension to other imaging contrasts and pathological cases. The online version contains supplementary material available at 10.1038/s41598-026-43804-1.

Approximately 30-50% of Papillary thyroid carcinoma (PTC) patients develop cervical lymph nodes (LNs) metastasis, significantly increasing the risk of disease recurrence and impacting long-term outcomes. We introduced an open-access multicenter lymph node ultrasound image database (LymphUs) specifically designed to advance research in LN assessment for PTC. Ultrasound imaging was performed on PTC patients at two independent clinical centers using standardized acquisition protocols. Experienced radiologists at each center documented sixteen semantic features for each LN. All LNs were annotated with segmentation masks serving as ground truth, and classification into benign or malignant categories was confirmed by fine needle aspiration biopsy results. The LymphUs comprises ultrasound images with segmentation masks from 338 PTC patients with suspected LN metastasis, divided into two center-specific cohorts: 180 patients (81 malignant, 99 benign) and 158 patients (82 malignant, 76 benign). The complete dataset, including semantic features and expert annotations, is freely accessible for research purposes. The LymphUs bridges a critical gap in medical imaging resources by providing a large-scale, multicenter ultrasound database for cervical LN assessment in PTC, supporting diagnostic algorithms, standardized reporting systems, and artificial intelligence applications to enhance preoperative LN staging and treatment planning.

Foundation models (FMs), such as the Segment Anything Model (SAM), have remarkable capabilities for general-purpose segmentation tasks through large-scale pre-training. However, a substantial domain shift limits their effectiveness in complex medical imaging. Here we introduce FM-Adapt, the first parameter-efficient adaptation of a FM (SAM-based vision transformer) into a resolution-agnostic architecture with photoacoustic (PA)-supervised learning for dual-target interventional ultrasound (US) segmentation. We demonstrate FM-Adapt in the context of PA-supervised interventions, specifically for US-guided needle tracking and simultaneous target identification (breast tumor segmentation). We train once with this unified adaptation framework to produce two specialized model weights: USPA-SAM for real-time tracking of needles and BT-SAM for segmenting breast tumors. This framework utilizes frozen pre-trained encoder components and fine-tunes only the mask decoder, allowing the model to process native (256 × 256) clinical images without spatial degradation while achieving state-of-the-art performance with high computational efficiency. USPA-SAM achieves a mean modified Hausdorff Distance (MHD) of 0.34 mm, a targeting error (TE) of 0.83 mm, and a 100% needle localization success rate (NLSR), outperforming baselines by a factor of 3- 17 × in spatial precision. Notably, on tumor segmentation, BT-SAM achieves Dice scores of 93.6% and 96.3%, along with IoU scores of 89.2% and 94.0%, demonstrating strong generalization to unseen data. This work demonstrates that our models achieve a 27 × improvement in computational efficiency to process native clinical images at 34 FPS on a single GPU to enable real-time clinical adaptation.

Recently, western countries have built evidence on mammographic artificial Intelligence-computer-aided diagnosis (AI-CADx) systems; however, their effectiveness has not yet been sufficiently validated in Japanese women. In this study, we aimed to establish a Japanese mammographic AI-CADx system for the first time. We retrospectively collected screening or diagnostic mammograms from 63 institutions in Japan. We then randomly divided the images into training, validation, and test datasets in a balanced ratio of 8:1:1 on a case-level basis. The gold standard of annotation for the AI-CADx system is mammographic findings based on pathologic references. The AI-CADx system was developed using SE-ResNet modules and a sliding window algorithm. A cut-off concentration gradient of the heatmap image was set at 15%. The AI-CADx system was considered accurate if it detected the presence of a malignant lesion in a breast cancer mammogram. The primary endpoint of the AI-CADx system was defined as a sensitivity and specificity of over 80% for breast cancer diagnosis in the test dataset. We collected 20,638 mammograms from 11,450 Japanese women with a median age of 55 years. The mammograms included 5019 breast cancer (24.3%), 5026 benign (24.4%), and 10,593 normal (51.3%) mammograms. In the test dataset of 2059 mammograms, the AI-CADx system achieved a sensitivity of 83.5% and a specificity of 84.7% for breast cancer diagnosis. The AUC in the test dataset was 0.841 (DeLong 95% CI; 0.822-0.859). The Accuracy was almost consistent independent of breast density, mammographic findings, type of cancer, and mammography vendors (AUC (range); 0.639-0.906). The developed Japanese mammographic AI-CADx system diagnosed breast cancer with a pre-specified sensitivity and specificity. We are planning a prospective study to validate the breast cancer diagnostic performance of Japanese physicians using this AI-CADx system as a second reader. UMIN, trial number UMIN000039009. Registered 26 December 2019, https://www.umin.ac.jp/ctr/.

In orthodontic treatments, accurately assessing the upper airway volume and morphology is essential for proper diagnosis and planning. Cone beam computed tomography (CBCT) is used for assessing upper airway volume through manual, semi-automatic, and automatic airway segmentation methods. This study evaluates upper airway segmentation accuracy by comparing the results of an automatic model and a semi-automatic method against the gold standard manual method. An automatic segmentation model was trained using the MONAI Label framework to segment the upper airway from CBCT images. An open-source program, ITK-SNAP, was used for semi-automatic segmentation. The accuracy of both methods was evaluated against manual segmentations. Evaluation metrics included Dice Similarity Coefficient (DSC), Precision, Recall, 95% Hausdorff Distance (HD), and volumetric differences. The automatic segmentation group averaged a DSC score of 0.915±0.041, while the semi-automatic group scored 0.940±0.021, indicating clinically acceptable accuracy for both methods. Analysis of the 95% HD revealed that semi-automatic segmentation (0.997±0.585) was more accurate and closer to manual segmentation than automatic segmentation (1.447±0.674). Volumetric comparisons revealed no statistically significant differences between automatic and manual segmentation for total, oropharyngeal, and velopharyngeal airway volumes. Similarly, no significant differences were noted between the semi-automatic and manual methods across these regions. It has been observed that both automatic and semi-automatic methods, which utilise open-source software, align effectively with manual segmentation. Implementing these methods can aid in decision-making by allowing faster and easier upper airway segmentation with comparable accuracy in orthodontic practice.

Checkpoint inhibitors targeting programmed cell death protein 1 (PD-1)/programmed death-ligand 1 (PD-L1) have revolutionized cancer therapy across many indications including urothelial carcinoma (UC). Because many patients do not benefit, a better understanding of the molecular mechanisms underlying response and resistance is needed to improve outcomes. We profiled tumors from 2,803 UC patients from four late-stage randomized clinical trials evaluating the PD-L1 inhibitor atezolizumab by RNA sequencing (RNA-seq), a targeted DNA panel, immunohistochemistry, and digital pathology. Machine learning identifies four transcriptional subtypes, representing luminal desert, stromal, immune, and basal tumors. Overall survival benefit from atezolizumab over standard-of-care is observed in immune and basal tumors, through different response mechanisms. A self-supervised digital pathology approach can classify molecular subtypes from H&E slides with high accuracy, which could accelerate tumor molecular profiling. This study represents a large integration of UC molecular and clinical data in randomized trials, paving the way for clinical studies tailoring treatment to specific molecular subtypes in UC and other indications.

Pathogenic activating mutations in the fibroblast growth factor receptor 3 (FGFR3) drive disease maintenance and progression in urothelial cancer. 10-15% of muscle-invasive and metastatic urothelial cancer (MIBC/mUC) are FGFR3-mutant. Selective targeting of FGFR3 hotspot mutations with tyrosine kinase inhibitors (e.g., erdafitinib) is approved for mUC and requires FGFR3 mutational testing. However, current testing assays (polymerase chain reaction or next-generation sequencing) necessitate high tissue quality, have long turnover time, and are expensive. To overcome these limitations, we develop a deep-learning model that detects FGFR3 mutations using routine hematoxylin-eosin slides. Encompassing 1222 cases, our study is a large-scale validation of a model prescreening FGFR3 mutations for MIBC and mUC patients. In this work, we demonstrate that our model achieves high sensitivity (>93%) on advanced and metastatic cases while reducing molecular testing by 40% on average, thereby offering a cost-effective and rapid pre-screening tool for identifying patients eligible for FGFR3 targeted therapies.

The digitization of health records and growing availability of tumour DNA sequencing provide an opportunity to study the determinants of cancer outcomes with unprecedented richness. Patient data are often stored in unstructured text and siloed datasets. Here we combine natural language processing annotations1,2 with structured medication, patient-reported demographic, tumour registry and tumour genomic data from 24,950 patients at Memorial Sloan Kettering Cancer Center to generate a clinicogenomic, harmonized oncologic real-world dataset (MSK-CHORD). MSK-CHORD includes data for non-small-cell lung (n = 7,809), breast (n = 5,368), colorectal (n = 5,543), prostate (n = 3,211) and pancreatic (n = 3,109) cancers and enables discovery of clinicogenomic relationships not apparent in smaller datasets. Leveraging MSK-CHORD to train machine learning models to predict overall survival, we find that models including features derived from natural language processing, such as sites of disease, outperform those based on genomic data or stage alone as tested by cross-validation and an external, multi-institution dataset. By annotating 705,241 radiology reports, MSK-CHORD also uncovers predictors of metastasis to specific organ sites, including a relationship between SETD2 mutation and lower metastatic potential in immunotherapy-treated lung adenocarcinoma corroborated in independent datasets. We demonstrate the feasibility of automated annotation from unstructured notes and its utility in predicting patient outcomes. The resulting data are provided as a public resource for real-world oncologic research.

Colorectal cancer is frequently diagnosed in advanced stages, highlighting the need for developing approaches for early detection. Liquid biopsy using cell-free DNA (cfDNA) fragmentomics is a promising approach, but the clinical application is hindered by complexity and cost. This study aimed to develop an integrated model using cfDNA fragmentomics for accurate, cost-effective early-stage colorectal cancer detection. Plasma cfDNA was extracted and sequenced from a training cohort of 360 participants, including 176 patients with colorectal cancer and 184 healthy controls. An ensemble stacked model comprising five machine learning models was employed to distinguish patients with colorectal cancer from healthy controls using five cfDNA fragmentomic features. The model was validated in an independent cohort of 236 participants (117 patients with colorectal cancer and 119 controls) and a prospective cohort of 242 participants (129 patients with colorectal cancer and 113 controls). The ensemble stacked model showed remarkable discriminatory power between patients with colorectal cancer and controls, outperforming all base models and achieving a high area under the receiver operating characteristic curve of 0.986 in the validation cohort. It reached 94.88% sensitivity and 98% specificity for detecting colorectal cancer in the validation cohort, with sensitivity increasing as the cancer progressed. The model also demonstrated consistently high accuracy in within-run and between-run tests and across various conditions in healthy individuals. In the prospective cohort, it achieved 91.47% sensitivity and 95.58% specificity. This integrated model capitalizes on the multiplex nature of cfDNA fragmentomics to achieve high sensitivity and robustness, offering significant promise for early colorectal cancer detection and broad patient benefit.  Significance: The development of a minimally invasive, efficient approach for early colorectal cancer detection using advanced machine learning to analyze cfDNA fragment patterns could expedite diagnosis and improve treatment outcomes for patients. See related commentary by Rolfo and Russo, p. 3128.

Probe-based confocal laser endomicroscopy (pCLE) enables real-time microscopic visualization of the duodenal mucosa and has shown acute food-triggered disruption of the duodenal epithelial barrier of patients with irritable bowel syndrome (IBS). The interpretation of the recordings is subjective, with unknown agreement rates. The aim of this study was to investigate the intra- and interobserver variability of this technique. An international multicenter study was performed, including pCLE recordings from three centers. Recordings were randomized and re-evaluated by five blinded experienced assessors. Low-quality recordings were excluded. The mucosa was considered altered if both fluorescein leakage and luminal particles were observed. Agreement was quantified using Fleiss' and Cohen's kappa (κ). Reference videos (i.e., videos with 100% agreement) were used to assess the optimal characteristics of videos needed to make a judgment based on the optimal receiver operating characteristic curve cutoff. Of the 119 individual recordings, 87 could be used for analyses (total of 86,408 frames). Intraindividual agreement rate was 80%-100%, whereas the interindividual agreement rate was 85% (κ = 0.68). The agreement rate with the endoscopist ranged 54%-95% (κ = 0.15-0.89). The optimal cutoff to distinguish altered from unaltered was by observing alterations in ≥ 2 out of 6 mucosal spots (100% sensitivity and specificity). Our study showed a substantial to perfect intraobserver agreement and a substantial interobserver agreement for the judgment of acute food-triggered disruption of the duodenal epithelial barrier by pCLE, confirming that this real-time readout is reliable and reproducible.

To develop an artificial intelligence (AI) model to diagnose Acanthamoeba keratitis (AK) based on in vivo confocal microscopy (IVCM) images extracted from the Heidelberg Retinal Tomograph 3 (HRT 3). This retrospective cohort study utilized HRT 3 IVCM images from patients who had received a culture-confirmed diagnosis of AK between 2013 and 2021 at Massachusetts Eye and Ear. Two cornea specialists independently labeled the images as AK or nonspecific finding (NSF) in a blind manner. Deep learning tasks were then conducted through Python and TensorFlow. Distinguishing between AK and NSF was designed as the task and completed through a devised convolutional neural network. A dataset of 3312 confocal images from 17 patients with a culture-confirmed diagnosis of AK was used in this study. The inter-rater agreement for identifying the presence or absence of AK in IVCM images was 84 %, corresponding to a total of 2782 images on which both observers agreed and which were included in the model. 1242 and 1265 images of AK and NSF, respectively, were utilized in the training and validation sets, and 173 and 102 images of AK and NSF, respectively, were utilized in the evaluation set. Our model had an accuracy, sensitivity, and specificity of 76 % each, and a precision of 78 %. We developed an HRT-based IVCM AI model for AK diagnosis utilizing culture-confirmed cases of AK. We achieved good accuracy in diagnosing AK and our model holds significant promise in the clinical application of AI in improving early AK diagnosis.

Diffuse midline glioma (DMG) is a rare, aggressive, and fatal tumor that largely occurs in the pediatric population. To improve outcomes, it is important to characterize DMGs, which can be performed via magnetic resonance imaging (MRI) assessment. Recently, artificial intelligence (AI) and advanced imaging have demonstrated their potential to improve the evaluation of various brain tumors, gleaning more information from imaging data than is possible without these methods. This narrative review compiles the existing literature on the intersection of MRI-based AI use and DMG tumors. The applications of AI in DMG revolve around classification and diagnosis, segmentation, radiogenomics, and prognosis/survival prediction. Currently published articles have utilized a wide spectrum of AI algorithms, from traditional machine learning and radiomics to neural networks. Challenges include the lack of cohorts of DMG patients with publicly available, multi-institutional, multimodal imaging and genomics datasets as well as the overall rarity of the disease. As an adjunct to AI, advanced MRI techniques, including diffusion-weighted imaging, perfusion-weighted imaging, and Magnetic Resonance Spectroscopy (MRS), as well as positron emission tomography (PET), provide additional insights into DMGs. Establishing AI models in conjunction with advanced imaging modalities has the potential to push clinical practice toward precision medicine.

Determining whether pediatric patients with low-grade gliomas (pLGGs) have tumor-related epilepsy (GAE) is a crucial aspect of preoperative evaluation. Therefore, we aim to propose an innovative, machine learning- and deep learning-based framework for the rapid and non-invasive preoperative assessment of GAE in pediatric patients using magnetic resonance imaging (MRI). In this study, we propose a novel radiomics-based approach that integrates tumor and peritumoral features extracted from preoperative multiparametric MRI scans to accurately and non-invasively predict the occurrence of tumor-related epilepsy in pediatric patients. Our study developed a multimodal MRI radiomics model to predict epilepsy in pLGGs patients, achieving an AUC of 0.969. The integration of multi-sequence MRI data significantly improved predictive performance, with Stochastic Gradient Descent (SGD) classifier showing robust results (sensitivity: 0.882, specificity: 0.956). Our model can accurately predict whether pLGGs patients have tumor-related epilepsy, which could guide surgical decision-making. Future studies should focus on similarly standardized preoperative evaluations in pediatric epilepsy centers to increase training data and enhance the generalizability of the model.

Amyloid PET imaging plays a crucial role in the diagnosis and research of Alzheimer's disease (AD), allowing non-invasive detection of amyloid-β plaques in the brain. However, the low spatial resolution of PET scans limits the accurate quantification of amyloid deposition due to partial volume effects (PVE). In this study, we propose a novel approach to addressing PVE using a latent diffusion model for resolution recovery (LDM-RR) of PET imaging. We leverage a synthetic data generation pipeline to create high-resolution PET digital phantoms for model training. The proposed LDM-RR model incorporates a weighted combination of L1, L2, and MS-SSIM losses at both noise and image scales to enhance MRI-guided reconstruction. We evaluated the model's performance in improving statistical power for detecting longitudinal changes and enhancing agreement between amyloid PET measurements from different tracers. The results demonstrate that the LDM-RR approach significantly improves PET quantification accuracy, reduces inter-tracer variability, and enhances the detection of subtle changes in amyloid deposition over time. We show that deep learning has the potential to improve PET quantification in AD, effectively contributing to the early detection and monitoring of disease progression.

Eye morphology varies significantly across the population, especially for the orbit and optic nerve. These variations limit the feasibility and robustness of generalizing population-wise features of eye organs to an unbiased spatial reference. To tackle these limitations, we propose a process for creating high-resolution unbiased eye atlases. First, to restore spatial details from scans with a low through-plane resolution compared with a high in-plane resolution, we apply a deep learning-based super-resolution algorithm. Then, we generate an initial unbiased reference with an iterative metric-based registration using a small portion of subject scans. We register the remaining scans to this template and refine the template using an unsupervised deep probabilistic approach that generates a more expansive deformation field to enhance the organ boundary alignment. We demonstrate this framework using magnetic resonance images across four different tissue contrasts, generating four atlases in separate spatial alignments. When refining the template with sufficient subjects, we find a significant improvement using the Wilcoxon signed-rank test in the average Dice score across four labeled regions compared with a standard registration framework consisting of rigid, affine, and deformable transformations. These results highlight the effective alignment of eye organs and boundaries using our proposed process. By combining super-resolution preprocessing and deep probabilistic models, we address the challenge of generating an eye atlas to serve as a standardized reference across a largely variable population.

"Just Accepted" papers have undergone full peer review and have been accepted for publication in Radiology: Artificial Intelligence. This article will undergo copyediting, layout, and proof review before it is published in its final version. Please note that during production of the final copyedited article, errors may be discovered which could affect the content. Background Detection and segmentation of lung tumors on CT scans are critical for monitoring cancer progression, evaluating treatment responses, and planning radiation therapy; however, manual delineation is labor-intensive and subject to physician variability. Purpose To develop and evaluate an ensemble deep learning model for automating identification and segmentation of lung tumors on CT scans. Materials and Methods A retrospective study was conducted between July 2019 and November 2024 using a large dataset of CT simulation scans and clinical lung tumor segmentations from radiotherapy plans. This dataset was used to train a 3D U-Net-based, image-multiresolution ensemble model to detect and segment lung tumors on CT scans. Model performance was evaluated on internal and external test sets composed of CT simulation scans and lung tumor segmentations from two affiliated medical centers, including single primary and metastatic lung tumors. Performance metrics included sensitivity, specificity, false positive rate, and Dice similarity coefficient (DSC). Model-predicted tumor volumes were compared with physician-delineated volumes. Group comparisons were made with Wilcoxon signed-rank test or one-way ANOVA. P < 0.05 indicated statistical significance. Results The model, trained on 1,504 CT scans with clinical lung tumor segmentations, achieved 92% sensitivity (92/100) and 82% specificity (41/50) in detecting lung tumors on the combined 150-CT scan test set. For a subset of 100 CT scans with a single lung tumor each, the model achieved a median model-physician DSC of 0.77 (IQR: 0.65-0.83) and an interphysician DSC of 0.80 (IQR: 0.72-0.86). Segmentation time was shorter for the model than for physicians (mean 76.6 vs. 166.1-187.7 seconds; p<0.001). Conclusion Routinely collected radiotherapy data were useful for model training. The key strengths of the model include a 3D U-Net ensemble approach for balancing volumetric context with resolution, robust tumor detection and segmentation performance, and the ability to generalize to an external site.

Tumour vascular density assessed from CD-31 immunohistochemistry (IHC) images has previously been shown to have prognostic value in breast cancer. Current methods to measure vascular density, however, are time-consuming, suffer from high inter-observer variability and are limited in describing the complex tumour vasculature morphometry. We propose a method for automatically measuring a range of vascular parameters from CD-31 IHC images, which together provide a detailed description of the vasculature morphology. We first used a U-Net based convolutional neural network, trained and validated using 36 partially annotated whole slide images from 27 patients, to segment vessel structures and tumour regions from which the measurements are taken. The model also segments the vascular smooth muscle, benign epithelium, adipose tissue, stroma, lymphocyte clusters, nerves and CD-31 positive leukocytes, and we applied it to an additional 21 images from 15 patients. Using these segmentations, we investigated the relationship between the various tissue types and the vasculature and studied the relationship of various vascular parameters with clinical parameters. We also performed a 3D histology analysis on a separate tumour sample as a proof of principle, providing a more comprehensive visualization of vasculature morphology compared to the standard 2D cross-section of a tissue sample. Using two-way cross-validation, we show that vessels were accurately segmented, with Dice scores of 0.875 and 0.856, and were accurately identified, with F1 scores of 0.777 and 0.748. All vascular parameters exhibit strong ( r > 0.7 ) and significant (p<0.001) correlations with measurements taken from the manual ground truth vessel segmentations. A significant relationship between the major/minor axis ratio, a measure of elongation, and the tumour grade was found. Our proposed method shows promise as a tool for studying the tumour vasculature and its relationship with surrounding cells and tissue types. Furthermore, the correlation with tumour grade highlights the clinical relevance of our approach. These findings suggest that our method could have substantial implications for improving prognostic assessments and personalizing therapeutic strategies in breast cancer treatment.

Emergency radiographic interpretation for fractures is prone to missed or misdiagnoses. Artificial intelligence (AI) is expected to become a powerful tool to assist clinicians in fracture detection. A systematic review and meta-analysis was performed to assess whether AI improves clinicians' ability to detect fractures on radiographs. A literature search was conducted in PubMed, Web of Science, and Cochrane Library for studies published between January 1, 2010, and October 10, 2025. A meta-analysis of diagnostic accuracy studies was performed using a Summary Receiver Operating Characteristic (SROC) curve. The quality of included studies was assessed using the Quality Assessment of Diagnostic Accuracy Studies 2 (QUADAS-2) tool. Subgroup analysis and meta-regression were conducted to explore potential sources of heterogeneity. A total of 26 studies were included . The pooled sensitivity of clinicians increased from 77% (95% CI: 72-81) to 87% (95% CI: 83-90) with AI assistance, while the pooled specificity improved from 88% (95% CI: 85-90) to 92% (95% CI: 89-94). The corresponding AUC values were 0.90 (95% CI: 0.87-0.92) before and 0.95 (95% CI: 0.93-0.97) after AI assistance. Eight studies were rated as high risk of bias. Subgroup analysis and meta-regression identified potential sources of heterogeneity, including fracture location, AI model type, high risk of bias, and reference standards. AI assistance significantly improves clinicians' diagnostic performance in detecting fractures on radiographs for extremity and trunk fractures.

Immunohistochemical (IHC) biomarkers such as estrogen receptor (ER), progesterone receptor (PR), HER2, and Ki-67 are essential for the classification and treatment of breast cancer. While radiomics-based models have demonstrated potential in non-invasive biomarker prediction, the utility of large language models (LLMs) for this task using only textual clinical data remains largely unexplored. This study aimed to evaluate the performance of ChatGPT-4o, a generative LLM, in predicting key IHC biomarkers based solely on structured radiological and pathological reports. Fifty-five patients with breast cancer were retrospectively analyzed. For each patient, structured clinical, imaging, and pathology reports-excluding IHC data-were entered into ChatGPT-4o. The model was prompted to generate predictions for ER, PR, HER2, and Ki-67 expression. Predictions were repeated at two time points to assess reproducibility. Diagnostic performance was compared to pathology results using accuracy, sensitivity, specificity, and Cohen's kappa. The model achieved the highest accuracy for HER2 prediction (83.6%, κ = 0.51), followed by ER (81.8%, κ = 0.44) and PR (76.4%, κ = 0.39). For high Ki-67 expression, the sensitivity was 88.9% with moderate overall agreement (κ = 0.55). Inter-prediction agreement was substantial to almost perfect for all biomarkers (κ = 0.69-0.83). ChatGPT-4o successfully predicted IHC biomarker status using only structured textual data. Its performance was comparable to radiomics models, offering a feasible and accessible AI tool to support early clinical decision-making, especially in resource-limited settings or before IHC results are available.

Inter-observer variability (IOV) in radiotherapy contouring remains a significant source of uncertainty, especially for complex anatomical regions. Autocontouring software, including both atlas-based and deep-learning-based models, aims to improve contouring consistency and reduce workload. A systematic review was conducted in accordance with PRISMA guidelines to evaluate the impact of autocontouring software on IOV. Twenty-five eligible studies were identified that quantitatively assessed IOV using these tools. Extracted data included anatomical site, observer and case numbers, contouring method and evaluation metrics. Most studies reported significant reductions in IOV with the use of autocontouring. Edited autocontours frequently achieved mean Dice Similarity Coefficient (DSC) values above 0.85 for clinical and planning target volumes and 0.90 for organs at risk (OARs) with well-defined anatomy, using deep learning methods. Deep-learning-based models demonstrated greater consistency compared to atlas-based methods. Structures such as the lungs, heart and bladder showed the most substantial improvements, while anatomically indistinct targets such as the prostate bed and pelvic lymph nodes showed limited benefit. However, discrepancies remained between observers for certain structures despite the use of automation. Overall, autocontouring tools, particularly deep-learning models, improve contouring consistency in radiotherapy planning. However, performance is strongly influenced by anatomical complexity and segmentation method. Larger multi-institutional studies and standardised evaluation protocols are needed to support widespread clinical adoption and strengthen quality assurance.

In the present paper, we compared the efficiency of six transfer learning models to detect malignant cells in urine cytology. We also applied an ensemble learning with weighted soft voting to assess its importance in the diagnostic accuracy in urine cytology. The voided urine samples of total 104 cases of urothelial cell carcinoma (UCC) and 86 cases with no malignancy (benign) were selected. All the 104 cases of UCC were histopathology proven high grade urothelial cell carcinoma (UCC). Urine with negative cytology reports were followed up clinically. There were 446 images of benign samples and 1369 images from malignant samples on 100 x objective. We applied six transfer learning models (DenseNet121, inception_v3, ResNet50, MobileNetV2, VGG16 and Xception) to detect malignant cells in urine. To compare the performance of different models, dynamic training optimization was done and each model was auto stopped after their maximum performance. In addition, an ensemble learning with soft voting was used including the top three models to enhance the diagnostic accuracy. Xception transfer model showed the highest sensitivity (88.57%), accuracy (86.55%), precision (80.52%) and FI score (84.35%). It was the best performing model. The other two top performing models were InceptionV3 and ResNet50. The area under curve of receiver operating characteristic (AUCROC) was ≤ 90 in all the transfer learning models. The accuracy, sensitivity, specificity, precision, F1-Score and AUCROC of the ensemble transfer learning model were as follows: 92.10%, 95.41%, 85.51%, 91.23%, 93.24% and 0.977 respectively. First time, we evaluated a large number of transfer learning models in urine cytology to detect malignant cells. All the models showed high sensitivity, specificity and accuracy. In addition, the ensemble learning technique with soft voting showed remarkable superior performance than individual top three transfer learning models. The techniques transfer learning and ensemble models have high potential to use in routine screening of urine.

Tubulin is a validated anticancer target, yet the clinical translation of colchicine-binding site inhibitors remains limited by toxicity and resistance. To accelerate the discovery of safer tubulin-targeting agents, we employed a machine learning (ML)-driven drug repurposing strategy integrating computational and experimental validation. Robust AutoQSAR classification models were trained on 279 curated tubulin inhibitors and used to screen 4500 US FDA-approved drugs, predicting 1800 compounds as potential tubulin inhibitors. These candidates were subjected to multistage structure-based virtual screening using Glide HTVS, SP, and XP docking, narrowing the selection from 698 (HTVS) and 350 (SP) to 38 compounds at the XP stage. Binding free-energy calculations (MM-GBSA) and 200 ns molecular dynamics simulations identified four stable colchicine-site binders: omeprazole, podofilox, sulfadoxine, and trimethoprim, exhibiting favourable binding energetics (Glide XP scores -10.06 to -8.12 kcal/mol; ΔG bind ranging from -10.06 to -8.12 kcal/mol; ΔG bind ranging from -64.16 to -38.61 kcal/mol). Biochemical tubulin polymerization assays confirmed tubulin inhibition, while cell-based cytotoxicity studies demonstrated low-micromolar antiproliferative activity of omeprazole and podofilox against melanoma (IC₅₀ = 4.32 ± 0.29 μM and 4.98 ± 0.37 μM, respectively) and colorectal cancer cells (IC₅₀ = 6.22 ± 0.22 μM and 5.76 ± 0.18 μM; n = 3). Overall, this study highlights a ML-guided drug repurposing framework that, unlike prior colchicine binding site-focused virtual screening studies, integrates large-scale ML prioritization with experimental validation to identify novel colchicine-site-targeted anticancer candidates.

Objective.Current automatic segmentation models in radiotherapy, which are predominantly unimodal and image-based, have limited generalizability due to boundary ambiguity and the lack of guideline integration. This study proposes a text-guided segmentation network, termed (TG-SegNet), for the automatic delineation of clinical target volumes (CTVs) in rectal cancer radiotherapy.Approach.Data from 567 preoperative patients with rectal cancer were retrospectively collected. Text prompts contained (i) patient case information (age, sex, tumor stage, tumor location, position) and (ii) guideline-derived descriptions indicating which CTV subsites should be included. TG-SegNet integrates computed tomography-derived visual features with structured clinical text prompts encoded by PubMedBERT, fused via cross-attention and fine-grained fusion. The model was trained on 452 patients and tested on 115. Its performance was compared with that of nnU-Net and two ablated variants (TG-SegNet without text prompts and TG-SegNet with simplified fusion). The evaluation comprised quantitative metrics, including dice similarity coefficient (DSC), 95% Hausdorff distance (HD95), mean surface distance (MSD), surface DSC (S-DSC), and average path length (APL), along with blinded expert scoring and an efficiency analysis. In additional analyses, we conducted text-prompt and module ablations.Main results.TG-SegNet achieved the best performance across all quantitative metrics: DSC 0.927 ± 0.022, HD95 7.01 ± 6.05 mm, MSD 1.94 ± 1.08 mm, S-DSC 0.799 ± 0.074, and APL 7372 ± 4452 (allp< 0.01). In clinical evaluation, TG-SegNet significantly improved target coverage, guideline adherence, and overall clinical acceptability compared with nnU-Net and ablations (p< 0.05), with boundary appropriateness comparable to nnU-Net. TG-SegNet had the shortest correction time (3.39 ± 1.10 min), corresponding to 82.1% time savings versus manual delineation. Text-prompt ablations suggested that the CTV-subsite prompt component contributed more to performance. Module ablations showed that both cross-attention and fine-grained fusion were beneficial.Significance.By integrating clinical semantics with imaging, TG-SegNet demonstrated superior accuracy, efficiency, and clinical acceptability over nnU-Net and ablated models, highlighting its potential for clinical translation.

Generative AI (GenAI) has advanced computational pathology through various image translation models. These models synthesize histopathological images from existing ones, facilitating tasks such as color normalization and virtual staining. Current models, while effective, are mostly dedicated to specific source-target domain pairs and lack scalability for multi-domain translations. Here, we introduce His-MMDM, a diffusion model-based framework enabling multi-domain and multi-omics histopathological image translation. His-MMDM is not only effective in performing existing tasks such as transforming cryosectioned images to FFPE ones and virtual immunohistochemical (IHC) staining but can also facilitate knowledge transfer between different tumor types and between primary and metastatic tumors. Additionally, it performs genomics- and/or transcriptomics-guided editing of histopathological images, illustrating the impact of driver mutations and oncogenic pathway alterations on tissue histopathology and educating pathologists to recognize them. These versatile capabilities position His-MMDM as a versatile tool in the GenAI toolkit for future pathologists.

To describe 1-year changes in the cornea as assessed by anterior segment optical coherence tomography (AS-OCT) and in vivo confocal microscopy (IVCM) for participants receiving a tissue graft generated from a new manufacturing process using cultivated autologous limbal epithelial cells. Cultivated autologous limbal epithelial cell grafts were produced in a 2-stage manufacturing process following a good manufacturing process-compliant protocol. AS-OCT and IVCM were completed at baseline and 12 months after the treatment in subsets of these participants. Secondary efficacy outcomes were determined based on improvement of central corneal epithelial morphology and thickness [corneal epithelial thickness (CET)] and presence of conjunctival or corneal cells in central cornea. Among 14 participants, 13 (93%) were male, 12 (86%) were white, the mean age was 46 ± 16 years. At baseline, CET was 53 (range: 34, 64) microns, and epithelial basal cell density was 3964 (range: 822-5788) cells/mm 2 ; the ratio of the cells at central cornea was 20% corneal and 90% conjunctival epithelial cells. At 12 months, the mean changes were 3 μm in CET ( P = 0.67), and 1967 cells/mm 2 in epithelial basal cell density ( P = 0.02); the proportion of the central cells improved to 75% corneal and 38% conjunctival epithelial cells. The AS-OCT and IVCM findings are consistent with the clinical improvement, indicating the reconstitution of the corneal phenotype and clearing of the optical axis. Nevertheless, IVCM is notably more effective for in-depth analysis of the epithelial phenotype and thickness than AS-OCT.

Magnetic resonance imaging (MRI) in children requires multiple sequences, leading to lengthy exams and motion-related challenges. Synthetic MRI generates multiple contrasts from a single acquisition, and integration of 3D-quantification using an interleaved Look-Locker acquisition sequence with a T2 preparation pulse (3D-QALAS) sequence with scan-specific deep-learning-based subspace reconstruction (Zero-DeepSub) enables high-resolution isotropic imaging with potential clinical utility. To evaluate synthetic images generated from a 3D-QALAS sequence with Zero-DeepSub relative to conventional MRI sequences in pediatric brain MRI. This prospective initial experience included 26 pediatric patients (mean age 8.4 years) who underwent clinically indicated brain MRI between November 2023 and January 2024. Synthetic T1-weighted, T2-weighted, and fluid-attenuated inversion recovery (FLAIR) images were generated from quantitative maps using 3D-QALAS with Zero-DeepSub reconstruction. Two neuroradiologists independently assessed seven predefined imaging findings on synthetic and conventional images, with discrepancies adjudicated by a third reader. This reader also performed a semiquantitative evaluation of image quality using a 5-point Likert scale. Statistical analysis included descriptive statistics, interobserver agreement (Cohen's kappa), and Wilcoxon signed-rank tests; positive predictive value (PPV) and negative predictive value (NPV) were also calculated. Synthetic images showed high sensitivity and specificity for mass/lesion, encephalomalacia, and collections, with perfect reader agreement. Gliosis demonstrated high sensitivity but moderate specificity for one reader. Abnormal enhancement had the lowest sensitivity (0.40). Interobserver agreement was moderate for gliosis (κ=0.55) and almost perfect (κ=0.83-1.00) for other findings. Semiquantitative evaluation revealed no significant difference between synthetic and conventional FLAIR, T1-weighted imaging, or post-contrast sequences (P>0.1), while conventional T2-weighted imaging was significantly superior (P<0.001). 3D-QALAS with Zero-DeepSub reconstruction enables the synthesis of high-resolution, clinically interpretable brain images in pediatric patients, including post-contrast sequences. While conventional T2-weighted imaging remained superior, other synthetic contrasts were rated comparable to conventional images. This promising technique holds potential to reduce scan times in pediatric neuroimaging protocols, but further optimization and validation are required before clinical implementation.

Artificial intelligence (AI) is increasingly being integrated into everyday tasks and work environments. However, its adoption in medical image analysis has progressed more slowly due to high clinical stakes, limited availability of labeled data, and substantial variability in imaging protocols and population. These challenges are further pronounced in the field of fetal, infant, and toddler (FIT) neuroimaging, where datasets are especially scarce and subject to large amounts of anatomical variability. However, deep learning (DL), a specific method within machine learning, which is itself a subfield of AI, has emerged as a powerful framework to adapt to the challenges of medical image analysis. This review is written for the broad FIT research community, including clinicians, neuroscientists, and develop mental scientists who may not have formal training in AI. To make the material accessible, we provide a concise overview of DL concepts before reviewing a selected, and non-exhaustive, list of applications of DL in FIT neuroimaging, including structural image analysis, enhancement of data acquisition, modeling of cognitive and perceptual processes, and automated video tagging. In closing, we discuss best practices for data curation, ongoing challenges, and opportunities for future research.

4D Flow Magnetic Resonance Imaging (4D Flow MRI) is a non-invasive technique for volumetric, time-resolved blood flow quantification. However, apparent trade-offs between acquisition time, image noise, and resolution limit clinical applicability. In particular, in regions of highly transient flow, coarse temporal resolution can hinder accurate capture of physiologically relevant flow variations. Deep learning-based post-processing techniques have shown promise in overcoming these issues using so-called super-resolution networks. However, while existing super-resolution research has primarily focused on spatial upsampling, temporal super-resolution remains largely unexplored. The aim of this study was therefore to implement and evaluate a residual data-driven network for temporal super-resolution 4D Flow MRI. To achieve this, an existing spatial network (4DFlowNet) was re-designed for temporal upsampling, adapting input dimensions, and optimizing internal layer structures. The model was trained and tested on synthetic 4D Flow MRI data derived from patient-specific in-silico models, followed by additional evaluation on clinically acquired in-vivo datasets. Overall, excellent performance was achieved with input velocities effectively denoised and temporally upsampled, with a mean absolute error (MAE) of 1.0 cm/s in an unseen in-silico setting, outperforming deterministic alternatives (linear interpolation MAE = 2.3 cm/s, sinc interpolation MAE = 2.6 cm/s). Further, the network synthesized high-resolution temporal information from unseen low-resolution in-vivo data, with strong correlation observed at peak flow frames. As such, our results highlight the potential of utilizing data-driven neural networks for temporal super-resolution 4D Flow MRI, enabling high-frame-rate flow quantification without extending acquisition times beyond clinically acceptable limits.

Accurate and timely identification of potentially necrotic intestinal segments during surgery is critical for surgical decision-making. However, existing classification methods heavily rely on labeled data or struggle to capture early subtle abnormal features, thereby limiting their accuracy and generalizability in clinical applications. This study aims to achieve sensitive detection of potentially necrotic intestinal segments through an unsupervised one-class classification method based on autoencoder (AE) residuals. The AE was trained using readily available multispectral data from the normal small intestine to construct residuals that amplify abnormal spectral differences. The residual features were used as input for three unsupervised one-class classification algorithms, namely Local Outlier Factor, Isolation Forest and Minimum Covariance Determinant (MCD) for detection. Validation data were collected from rabbit models under different occlusion durations. The construction of residuals significantly improved the performance of all classifiers. In the potentially necrotic phase at 10 min of occlusion, the overall accuracy of the algorithms was steadily improved at about 80%, especially the sensitivity of MCD jumped from 26.5% to 80.7%, ensuring high detection rates from the outset. When the occlusion duration was extended, the sensitivity of all algorithms showed a positive correlation, consistent with the object physiological patterns. Final visualization results confirmed that residual features improved classification performance and preserved spatial continuity. This approach shows promise for early anomaly detection in other medical imaging applications.

Accurate volume measurements of subcortical brain structures are essential in neuroimaging studies. This study aimed to compare two automated segmentation algorithms, 3D Slicer and VolBrain, for subcortical brain volumetry. Subcortical volume measurements obtained using 3D Slicer and VolBrain were compared using the Neurofeedback Skull-stripped (NFBS) dataset. Statistical analyses included the coefficient of determination (R²) and Bland-Altman (B-A) plots. Volumetric measurements from both tools were evaluated across multiple subcortical brain regions. Strong agreement was observed for the caudate nucleus, thalamus, and hippocampus, while notable discrepancies were identified in the amygdala, globus pallidus, and nucleus accumbens. Both constant and proportional systematic errors were detected in certain regions, with greater variability observed in 3D Slicer measurements. Intraclass Correlation Coefficient (ICC) analysis demonstrated weak agreement for the accumbens and amygdala, and moderate agreement for the hippocampus and thalamus, highlighting region-specific differences in performance Conclusion:Both VolBrain and 3D Slicer are effective tools for subcortical brain volumetry. VolBrain provides a fully automated workflow, whereas 3D Slicer allows greater flexibility through detailed corrections. Method selection should be guided by research objectives, regions of interest, and image quality. Further validation across diverse datasets is warranted to improve clinical applicability, particularly in regions showing lower agreement.

Phosphocreatine (PCr) is an essential marker of muscle metabolism, and accurate quantification of its (fs) and its exchange rate (ksw) is essential for diagnosing various muscular and neuromuscular diseases. Although chemical exchange saturation transfer (CEST) MRI can detect the saturation transfer effect from PCr, quantification of the underlying PCr fs and ksw, particularly at low fields, remains challenging due to significant overlapping confounding effects in tissues when using conventional fitting approaches. Deep learning (DL) presents a promising alternative, yet traditional DL models often struggle to capture subtle PCr-specific variations induced by changes in fs or ksw. Furthermore, these models are typically trained on either fully synthetic data, which may not adequately mimic tissues, or in vivo data which lack ground truth. This study introduces a global-local two-branch DL model to effectively eliminate confounding effects and capture subtle variations in the PCr CEST effect. Furthermore, our model was trained on partially synthetic data that offers both simulation flexibility and fidelity. Model accuracy was evaluated by using both digital and physical phantoms, and the model was applied to skeletal muscle of healthy rats and rats with amyotrophic lateral sclerosis (ALS). Phantom experiments demonstrate that our approach surpasses all fitting methods, the state-of-the-art model, and other combinations of DL models and training data. In vivo, the model identified a significant reduction in PCr fs in ALS rats, which other methods fail to detect. Our global-local two-branch DL model trained using partially synthetic data enhances PCr quantification in skeletal muscle.

Liver MRI plays a critical role in the diagnosis and monitoring of liver disease; however, image quality is often degraded by respiratory motion and noise, particularly in high-resolution and diffusion-weighted imaging. Propeller-based sequences such as BLADE improve motion robustness, but advanced reconstruction strategies are required to fully exploit their potential under accelerated conditions. We propose a phase-constrained zero-shot self-supervised learning (PC ZS-SSL) framework for BLADE liver MRI reconstruction. The method embeds BLADE forward and adjoint operators within an unrolled deep network together with an explicit phase estimation module. Unlike supervised approaches, PC ZS-SSL requires no external training data and instead leverages partitioned k-space. The framework was first evaluated using structured phantom experiments assessing spatial resolution and low-contrast detectability, and subsequently applied to in vivo 3 T T2-weighted (T2W, 20 blades) and diffusion-weighted imaging (DWI, 18 and 15 blades, b = 0, 800 s/mm2). Performance was compared with locally low-rank (LLR) and vendor-provided reconstructions. In phantom studies, PC ZS-SSL preserved fine structural details and demonstrated improved sharpness compared with fewer gradient updates and LLR. In vivo, PC ZS-SSL consistently reduced noise and ringing artifacts while maintaining anatomical fidelity. In T2W imaging, it achieved image quality comparable to LLR with fewer artifacts. In DWI, where noise is more pronounced, PC ZS-SSL provided clearer organ boundaries than both LLR and vendor reconstructions. Additional in vivo evaluation demonstrated that PC ZS-SSL remained robust under severe undersampling conditions (e.g., 4 blades) and challenging imaging scenarios, where LLR reconstructions exhibited residual artifacts. PC ZS-SSL enables high-quality, artifact-suppressed BLADE liver MRI without the need for external training data. Its strong performance in both phantom and in vivo experiments-particularly under high-noise diffusion conditions-highlights its potential for clinical translation.

Bladder cancer (BLCA) is one of the most common urological malignancies, entailing significant morbidity and mortality rates. A comprehensive understanding of the mechanisms driving bladder cancer initiation and development is required to devise improved treatment regimens. N6-methyladenosine (m6A) is the most prevalent internal RNA modification reported to regulate cancer metastasis. Here, the function of fat mass and obesity-associated protein (FTO), an m6A demethylase, was studied in bladder cancer by overexpressing or knocking down FTO, and the underlying mechanism of FTO in BLCA was explored using machine learning analysis, in vitro, and in vivo experiments. The FTO expression has been proven to be downregulated within bladder cancer, and it could exert a tumor-suppressive effect. Moreover, Gain- and loss-of-function experiments showed that FTO downregulation enhanced the proliferation, migration, and invasion of bladder cancer cells. Mechanistic studies revealed that FTO decreased EMG1 expression by demethylating EMG1 and reducing ribosome biosynthesis, thereby promoting bladder cancer cell proliferation, migration, and invasion and repressing tumor growth in vivo. This study has demonstrated the anti-tumor effect of FTO on bladder cancer development, making it a promising therapeutic target.

Adoptive cell therapy (ACT) with tumor-infiltrating lymphocytes (TIL) is a form of personalized immunotherapy that requires ex vivo expansion of autologous TILs and their reinfusion back into the patient. Predicting TIL expansion at the time of diagnosis may improve selection of patients that can benefit from ACT-TIL. It can also prevent high treatment-related costs and delays in treatment of patients whose cancer specimens would not yield successful TIL growth. We developed PETIL, a machine-learning model optimized for data of a medium size to determine a minimal combination of features (demographic, clinical, and biological specimen-based) that is predictive of expansion of TILs from a resected bladder cancer. We used a retrospectively identified set of data from bladder cancer patients at Moffitt Cancer Center for the training and testing cohorts. Additionally, we used data from a recent feasibility clinical trial at Moffitt Cancer Center as a blinded validation cohort. PETIL uses random forest method to identify a combination of robust predictive features, support vector machine model to determine the optimal classification hyperparameters, and Matthews correlation coefficient method to adjust the decision-boundary threshold for imbalanced data. Our model yielded AUC=0.740 for the testing cohort and AUC=0.857 for blinded validation cohort. Thus, our PETIL model optimized for data of medium size has favorable performance metrics for predicting TIL expansion from a given tumor. Treatment with autologous tumor-infiltrating lymphocytes (TIL) that are expanded ex vivo from a given tumor and then reinfused into the patient is a promising personalized immunotherapy. However, the TIL expansion takes about 4-6 weeks, thus developing tools that predict whether TIL growth will be successful can help to avoid delays in treatment of patients whose cancer specimens would not yield successful TIL expansion. Our Predictor of Expansion of TIL (PETIL) is a machine-learning model that uses patients' demographic information, clinical tumor classification, and biological tumor specimen-based measurements to determine a minimal set of these data features that are predictive of TIL expansion outcome. We applied this model to data from bladder cancer patients collected at Moffitt Cancer Center and showed that PETIL has favorable performance metrics for the dataset of a moderate size. This computational predictor can support clinicians in determining which patients are candidates for TIL immunotherapy. The developed PETIL pipeline can also be adjusted to data from other solid tumors.

To develop and validate a robust, multimodal machine learning framework integrating radiomic and deep learning features from multiplex immunohistochemistry (mIHC) images for comprehensive outcome prediction in colorectal cancer (CRC). This multi-institutional retrospective study included 2,117 CRC patients from seven centers, with 1,548 cases used for model training and internal testing, and 569 for external validation. mIHC-stained whole-slide images targeting six immune markers (CD3, CD8, CD45RO, PD-1, LAG-3, Tim-3) were analyzed from two spatial compartments: tumor center and invasive margin. Radiomic features (n = 71/region/marker) were extracted using HistomicsTK, while 768-dimensional deep features were derived using a pre-trained Vision Transformer (ViT-B/16). Feature robustness across biomarkers was quantified via intraclass correlation coefficients (ICC ≥ 0.75). Selected features underwent multi-step selection (LASSO, MI, RFE) and were fused into a single feature space, followed by PCA-based dimensionality reduction. Five clinical tasks were modeled: tumor recurrence, survival status, overall survival duration, TNM staging, and immune profile classification. Classification models (TabTransformer, XGBoost, TabNet) and survival models (DeepSurv, CoxPH, RSF) were trained using 5-fold cross-validation and tested on independent cohorts. Fused features significantly outperformed individual modalities across all tasks. TabTransformer with LASSO-selected fused features achieved top performance: recurrence (AUC = 95.9%), survival status (AUC = 94.5%), TNM staging (macro-AUC = 91.0%), and immune profile (macro-AUC = 91.0%). For survival regression, DeepSurv achieved a C-index of 0.82 and time-dependent AUC of 0.82. Models exhibited strong generalizability, with negligible performance drop on external datasets. SHAP analysis confirmed feature interpretability, with fused features contributing the most across tasks. This study demonstrates that fused mIHC-derived radiomic and deep features yield accurate, interpretable, and generalizable predictions for multiple CRC outcomes, supporting their integration into precision oncology workflows.

Pelvimetry may aid preoperative planning in rectal cancer surgery, yet manual measurements are time-consuming and MRI-based methods require dedicated protocols. We developed a fully automated CT-based pipeline for quantifying mid-pelvic geometry and soft tissue occupancy at the ischial spine level, the narrowest corridor encountered during total mesorectal excision. This retrospective feasibility study included 73 patients with mid-to-low rectal cancer who underwent contrast-enhanced staging CT. Automated segmentation was performed using TotalSegmentator. The pipeline extracted interspinous distance (ISD) via an anatomically anchored search strategy with valley detection, constructed a posterior pelvic triangle bounded by the bilateral ischial spines and anterior sacrum, and quantified bowel and fat occupancy within this region. Automated measurements were validated against blinded manual annotations by two independent raters. Pipeline success rates, failure taxonomy, and agreement between contrast-enhanced and non-contrast acquisitions (n = 69 paired cases) were evaluated using intraclass correlation coefficients (ICC) and Bland-Altman analysis. The pipeline achieved complete ISD extraction in 100% of contrast-enhanced and 95.8% of non-contrast cases. In blinded validation (n = 70), automated ISD agreement against manual reference (ICC = 0.977; bias = 0.45 mm) exceeded inter-rater reliability (ICC = 0.962), and triangle-derived metrics showed a good-to-excellent agreement (ICC = 0.86-0.89, auto vs. manual). ISD measurements showed excellent agreement across contrast conditions (ICC = 0.99; mean bias = 0.09 mm; 95% limits of agreement: - 2.51 to 2.69 mm). Bone-derived triangle metrics demonstrated strong concordance (ICC = 0.90-0.93). Fat-related metrics showed systematic differences between contrast conditions but maintained good agreement (ICC = 0.87-0.91). Sex-based differences were consistent with known pelvic dimorphism. This fully automated pipeline reliably extracts mid-pelvic geometry and soft tissue metrics from routine staging CT, with measurement agreement matching or exceeding manual inter-rater variability, offering a standardized, practical, and reproducible approach for characterizing the surgical workspace in rectal cancer patients.

Artificial intelligence (AI) could help medical practitioners in analyzing radiological images to determine the presence and site of bowel obstruction. This retrospective diagnostic study proposed a series of deep learning (DL) models for diagnosing bowel obstruction on abdominal radiograph. A total of 2082 upright plain abdominal radiographs were retrospectively collected from four hospitals. The images were labeled as normal, small bowel obstruction and large bowel obstruction by three senior radiologists based on comprehensive examinations and interventions within 48 hours after admission. Gradient-weighted class activation mapping was used to visualize the inferential explanation. In the validation set, the Xception-backboned model achieved the highest accuracy (0.863), surpassing the VGG16 (0.847) and ResNet models (0.836). In the test set, the Xception model (accuracy: 0.807) outperformed other models and a junior radiologist (0.780) but not a senior radiologist (0.840). In the AI-aided diagnostic framework, the junior and senior radiologists made their judgements while aware of the Xception model predictions. Their accuracy significantly improved to 0.887 and 0.913, respectively. We developed and validated DL-based computer vision models for diagnosing bowel obstruction on plain abdominal radiograph. DL-based computer-aided diagnostic systems could reduce medical practitioners' workloads and improve diagnostic accuracy.

The current pathological diagnosis of lymph node metastasis is time-consuming, labor-intensive, and dependent on sectioning of paraffin blocks. Herein, in a prospective cohort of patients with breast cancer, we validated dynamic full-field optical coherence tomography (D-FFOCT), a virtual pathology tool integrating deep learning for nodal metastasis detection, and offering rapid and label-free histologic approximations of fresh tissues. In a prospective dual-center cohort of 155 patients with breast cancer, 747 freshly bisected lymph node slides were obtained via D-FFOCT. Surgeons interpreted each slide with histopathology as the gold standard. A deep learning model was trained on 28,911 patches (corresponding to 590 slides) and tested on 7,736 patches (corresponding to 157 slides). The results were mapped to the slide level for potential intraoperative evaluation. D-FFOCT strongly correlated with hematoxylin and eosin (H&E)-stained histological images. Surgeons achieved 97.10% specificity in nodal diagnosis with D-FFOCT. The performance of the artificial intelligence (AI) model was not inferior to that of human experts and had a sensitivity/specificity of 87.88%/91.94% and an area under the receiver operating characteristic curve of 0.899 at the slide level. The human-AI collaborative system reduced labor requirements by 75% and increased the specificity by 6.5%, to 98.39%. D-FFOCT has excellent potential as a tool for assessing lymph node metastatic status without tissue preparation or consumption. The integration of D-FFOCT with deep learning decreases labor demands and maintains high accuracy, thereby enabling streamlined nodal prediction independent of routine pathology procedures.

Deep learning-based virtual fiber staining provides a promising complement to routine H&E pathology. However, the reliance on predefined staining style inputs and manual intervention limits the clinical applicability of existing methods. To address these challenges, we introduce ViFIT-assisted histopathology, a two-stage diagnostic approach that integrates our proposed unsupervised deep learning-based virtual fiber transformation model (ViFIT). This approach enables the conversion of H&E-stained images with diverse styles into pathologist-preferred H&E images, while simultaneously generating content-consistent virtual fiber images containing label-free collagen fibers and stained reticular and elastic fibers. ViFIT-assisted histopathology reveals tumor-associated fiber structures and provides quantitative metrics across multiple intraoperative and postoperative cases. Experimental results demonstrate that ViFIT significantly outperforms state-of-the-art unsupervised methods in both style standardization and virtual staining, across various downstream tasks and cancer types. By eliminating the need for staining variation and manual annotation, ViFIT-assisted histopathology streamlines histopathology workflows, making it well-suited for multi-center consultations and differential diagnosis.

Artificial intelligence (AI) and radiomics are increasingly applied in pediatric neuroradiology to enhance diagnostic precision. However, their clinical implementation remains limited due to methodological variability and lack of standardization. To systematically evaluate the diagnostic applications, performance, and methodological quality of artificial intelligence models, including both radiomics/machine learning and deep learning approaches, in pediatric brain tumor imaging. A systematic review was conducted in accordance with PRISMA 2020 guidelines, searching PubMed, Scopus, and Web of Science (2010-2025). Eligible studies included patients aged 0-18 years with brain tumors, applied AI to neuroimaging for diagnostic classification, molecular characterization, or integral image processing tasks, and reported performance metrics. Methodological quality was assessed using the Newcastle-Ottawa Scale (NOS) and the Checklist for Artificial Intelligence in Medical Imaging (CLAIM). From 638 records, 24 studies were included. Most studies used MRI (96%) and machine learning models based on radiomics (88%), with a smaller proportion employing deep learning (29%). The primary diagnostic tasks were tumor classification (50%) and molecular subtype prediction (33%). Reported AUCs for diagnostic tasks ranged from 0.73 to 0.98 (median: 0.91). Based on the NOS, 19 studies (79%) were rated as low risk of bias (scores 8-9), though only 3 studies (12.5%) performed external validation on independent cohorts. AI and radiomics demonstrate high diagnostic accuracy for pediatric brain tumor characterization. Nonetheless, the lack of prospective design and critically low rate of external validation limits generalizability. Standardized, multicenter studies are needed to support broader clinical adoption.

Brain tumors are the most common solid malignancies in children, encompassing diverse histological, molecular subtypes, and imaging features and outcomes. Pediatric brain tumors (PBTs), including high- and low-grade gliomas (HGG and LGG), medulloblastomas (MB), ependymomas, and rarer forms, pose diagnostic and therapeutic challenges. Deep learning (DL)-based segmentation offers promising tools for tumor delineation, yet its performance across heterogeneous PBT subtypes and MRI protocols remains uncertain. A retrospective single-center cohort of 174 pediatric patients with HGG, LGG, MB, ependymomas, and other rarer subtypes was used. MRI sequences included T1, T1 post-contrast (T1-C), T2, and FLAIR. Manual annotations were provided for 4 tumor subregions: whole tumor (WT), T2-hyperintensity (T2H), enhancing tumor (ET), and cystic component (CC). A 3D nnU-Net model was trained and tested (121/53 split), with segmentation performance assessed using the Dice similarity coefficient (DSC) and compared against intra- and interrater variability. The model achieved robust performance for WT and T2H (mean DSC: 0.85), comparable to human annotator variability (mean DSC: 0.86). ET segmentation was moderately accurate (mean DSC: 0.75), while CC performance was poor (mean DSC: 0.26). Segmentation accuracy varied by tumor type, MRI sequence combination, and location. Notably, T1, T1-C, and T2 combined produced results nearly equivalent to the full protocol. DL-based segmentation is feasible for PBTs, particularly for T2H and WT. Challenges remain for ET and CC segmentation, highlighting the need for further refinement. These findings support the potential for protocol simplification and automation to enhance volumetric assessment and streamline pediatric neuro-oncology workflows.

Amyloid-related imaging abnormalities with hemorrhage (ARIA-H) are a key safety concern in anti-amyloid therapies for Alzheimer's disease, as they are radiologically indistinguishable from cerebral microbleeds (CMBs). Accurate detection of CMBs is therefore essential for both treatment eligibility assessment and post-treatment safety monitoring. However, manual identification on 2D T2*-weighted gradient-recalled echo (GRE) MRI is labor-intensive and subject to variability. To develop and validate an artificial intelligence (AI)-based model for automated CMB detection using only 2D T2*-weighted GRE MRI, which is widely used in clinical settings. We implemented a YOLOv11-based deep learning model, preceded by a novel multi-channel preprocessing pipeline that enhances CMB visibility. The model was trained and tested using a dataset of 758 participants, with expert consensus used as the reference standard. Using the optimized basic preprocessing with super-resolution (BP + SR) pipeline, the model achieved a lesion-level sensitivity of 0.694, precision of 0.705, and F1-score of 0.699. In patient-level analysis for detecting elevated CMB burden (≥4), the system demonstrated sensitivity of 0.933 and specificity of 0.935, supporting reliable stratification of CMB severity. Regional analysis showed sensitivity of 0.625 for lobar CMBs and 0.627 for deep structures. This study demonstrates the feasibility of robust CMB detection using only 2D T2*-weighted GRE MRI. Based on current performance, we position this system as a decision-support tool for GRE-based CMB screening, in which lesion-level detections may be aggregated to inform patient-level CMB burden relevant to ARIA-H risk stratification, while final ARIA grading and clinical decisions require expert neuroradiological confirmation.

Magnetic resonance imaging (MRI) super-resolution aims to enhance spatial resolution from low-resolution acquisitions while preserving anatomically meaningful structures. Conventional single-image super-resolution methods are fundamentally limited by the lack of high-frequency information in the input and often fail to recover fine anatomical details under large upscaling factors. In many MRI scenarios, additional reference images acquired under similar imaging protocols are naturally available and can provide complementary structural information, yet effectively leveraging such references without introducing anatomically inconsistent artifacts remains challenging. In this work, we propose a reference-guided MRI super-resolution framework, termed Dual Attention Aggregation Super-Resolution (DAASR), which explicitly balances structural modeling and controlled reference integration in a hierarchical manner. DAASR employs a channel-wise attention mechanism to reinforce global anatomical coherence in low-resolution features and a structure-aware alignment strategy to selectively incorporate consistent reference information while suppressing unreliable transfers. Extensive experiments on a public MRI benchmark and a clinical brain MRI dataset demonstrate that the proposed method consistently outperforms state-of-the-art MRI super-resolution approaches across multiple scaling factors in terms of PSNR and SSIM. The proposed DAASR further achieves improved structural similarity and visual consistency, indicating better preservation of fine anatomical structures and tissue boundaries.

To explore a deep learning (DL) model for determining the relationship between the maxillary sinus (MS) and maxillary posterior teeth (MPT) based on cone beam computed tomography (CBCT) images and measuring the distance automatically between the MS and MPT using a 3D point cloud algorithm. A CBCT dataset containing 88 maxillary sinuses (MSs) and 352 maxillary posterior teeth (MPT) was annotated, and the MS-MPT distances were measured by clinicians as the ground truth. A segmentation model for MSs and MPT in CBCT images based on the U-Net convolutional block attention (CBAM) architecture was trained and assessed using a 3-fold cross-validation strategy. Then, calibrated point clouds were reconstructed using segmented anatomical structure data, and the Euclidean distances between the MS and MPT were measured; the minimum distance was identified as the MS-MPT distance. The performance of the model in terms of segmentation and distance measurement was evaluated, and the results were compared with the ground truth. Our segmentation model achieved a mean Dice similarity coefficient (DSC) of 0.959 and a mean Jaccard coefficient of 0.922 for MSs and a mean DSC of 0.913 and a mean Jaccard coefficient of 0.851 for MPT. The MS-MPT distances determined by clinicians and the 3D point cloud method demonstrated strong consistency (ϒ > 0.993, p < 0.01). In terms of the model and clinicians, the mean negative signed error was 0.63 mm (95% CI, 0.59-0.66 mm), and the successful detection rate (SDR) for the root apex of MPT reached 70.3% at the 1 mm threshold. In this study, an automated framework that combines deep learning-driven segmentation and three-dimensional point cloud analysis was developed to quantify the relationship between the maxillary sinus and maxillary posterior teeth and achieved reliable detection accuracy across diverse anatomical variations in CBCT scans.

The development of robust algorithms for lung and lobe segmentation is essential for diagnosing and monitoring pulmonary diseases. Obtaining manual or automatic annotations is challenging, especially in patients with severe abnormalities due to poorly visible lobar fissures. We present LobePrior, an automated lung lobe segmentation method combining deep neural networks and probabilistic models. Segmentation occurs in three stages: a coarse stage processing downsampled images, a high-resolution stage where specialized AttUNets segment each lobe, and a final post-processing stage. Probabilistic models derived from label fusion guide the network in regions with severe abnormalities, and synthetic lesion generation provides augmentation during training. Performance was evaluated on LOLA11 and three additional datasets with cancerous nodules or COVID-19 consolidations. LobePrior achieved accurate segmentations compared to manual ground truth, reaching state-of-the-art performance even in challenging cases. On the LOCCA dataset, it obtained a Dice score of 0.966, with similar improvements on a COVID-19 CT dataset (Dice 0.978). Statistically significant improvements over competing methods were observed across all datasets. These results demonstrate that LobePrior effectively integrates anatomical priors and deep learning to provide reliable lobe segmentation in the presence of severe pulmonary abnormalities.

To train and validate segmentation models for automated segmentation of gallbladder cancer (GBC) lesions from contrast-enhanced CT images. This retrospective study comprised consecutive patients with pathologically proven treatment naïve GBC who underwent a contrast-enhanced CT scan at four different tertiary care referral hospitals. The training and validation cohort comprised CT scans of 317 patients (center 1). The internal test cohort comprised a temporally independent cohort (n = 29) from center 1 (internal test 1). The external test cohort comprised CT scans from three centers [ (n = 85)]. We trained the state-of-the-art 2D and 3D image segmentation models, SAM Adapter, MedSAM, 3D TransUNet, SAM-Med3D, and 3D-nnU-Net, for automated segmentation of the GBC. The models' performance for GBC segmentation on the test datasets was assessed via dice score and intersection over union (IoU) using manual segmentation as the reference standard. The 2D models performed better than 3D models. Overall, MedSAM achieved the highest dice and IoU scores on both the internal [mean dice (SD) 0.776 (0.106) and mean IoU 0.653 (0.133)] and external [mean dice (SD) 0.763 (0.098) and mean IoU 0.637 (0.116)] test sets. Among the 3D models, TransUNet showed the best segmentation performance with mean dice (SD) and IoU (SD) of 0.479 (0.268) and 0.356 (0.235) in the internal test and 0.409 (0.339) and 0.317 (0.283) in the external test sets. The segmentation performance was not associated with GBC morphology. There was weak correlation between the dice/IoU and the size of the GBC lesions for any segmentation model. We trained 2D and 3D GBC segmentation models on a large dataset and validated these models on external datasets. MedSAM, a 2D prompt-based foundational model, achieved the best segmentation performance.

Acute pulmonary embolism (APE) is a common pulmonary condition that, in severe cases, can progress to right ventricular hypertrophy and failure, making it a critical health concern surpassed in severity only by myocardial infarction and sudden death. CT pulmonary angiogram (CTPA) is a standard diagnostic tool for detecting APE. However, for treatment planning and prognosis of patient outcome, an accurate assessment of individual APEs is required. Within this study, we compiled and prepared a dataset of 200 CTPA image volumes of patients with APE. We then adapted two state-of-the-art neural networks; the nnU-Net and the transformer-based VT-UNet in order to provide fully automatic APE segmentations. The nnU-Net demonstrated robust performance, achieving an average Dice similarity coefficient (DSC) of 88.25 ± 10.19% and an average 95th percentile Hausdorff distance (HD95) of 10.57 ± 34.56 mm across the validation sets in a five-fold cross-validation framework. In comparison, the VT-UNet was achieving on par accuracies with an average DSC of 87.90 ± 10.94% and a mean HD95 of 10.77 ± 34.19 mm. We applied two state-of-the-art networks for automatic APE segmentation to our compiled CTPA dataset and achieved superior experimental results compared to the current state of the art. In clinical routine, accurate APE segmentations can be used for enhanced patient prognosis and treatment planning.

The Coiled-coil domain-containing (CCDC) family, due to its unique protein structural domain and broad involvement in diverse biological processes, has emerged as a focus in oncology research. Nevertheless, its clinical significance and function in bladder cancer (BLCA) remain poorly defined. Machine learning algorithms were employed to identify pivotal CCDC genes in the cancer genome atlas (TCGA), and a prognostic model was subsequently constructed. Multi-omics data encompassing pan-cancer cohorts, single-cell sequencing, and spatial transcriptomics were integrated to characterize the expression patterns and prognostic significance of Coiled-coil domain-containing 137 (CCDC137), a previously uncharacterized CCDC family member in BLCA. Tissue microarray confirmed CCDC137 abnormal expression in bladder carcinoma specimens. The effect of CCDC137 knockdown on BLCA progression was evaluated through CCK8 assay, clonogenic formation, wound healing, Transwell, and subcutaneous xenograft models. RNA sequencing, quantitative RT-PCR, and western blot were utilized to delineate its regulatory network. A prognostic model incorporating 10 CCDC genes was successfully established in the TCGA-BLCA cohort. Then, we found that CCDC137 exhibited pan-cancer overexpression and usually correlation with poor clinical outcomes. Immunohistochemistry further substantiated its dysregulation in bladder carcinoma. Integrated multi-omics analyses suggested associations between CCDC137 expression and a tumor immunosuppressive microenvironment. CCDC137 knockdown significantly suppressed bladder cancer cell proliferation and migratory capacity in vitro. Correspondingly, subcutaneous xenograft tumor growth was inhibited in vivo. Moreover, decreased expression of stearoyl-CoA desaturase (SCD), a key lipid metabolic enzyme, accompanied CCDC137 depletion. These findings collectively suggest a cancer-promoting role for CCDC137 in bladder carcinoma. This systematic investigation combining multi-omics bioinformatics analyses and experimental validation demonstrates the role of CCDC137 in bladder carcinoma progression, providing novel mechanistic insights into the pathogenesis of BLCA and offering a theoretical foundation for therapeutic targeting of CCDC137 in urothelial malignancies.

Bladder cancer (BLCA) is among the most common malignancies worldwide, with significant mortality rates. The function of PANoptosis in BLCA, as a controlled process of programmed cell death, remains largely unelucidated. The study aimed to elucidate the role of PANoptosis-related genes in BLCA and investigate their molecular mechanisms, prognostic significance, and therapeutic potential. By analyzing differentially expressed genes in BLCA from The Cancer Genome Atlas (TCGA) and PANoptosis-associated genes, we discovered 98 genes associated with PANoptosis. Functional enrichment and consensus clustering identified molecular subtypes linked to these genes. A prognostic model was developed via LASSO regression based on these genes. Subsequent analyses assessed clinical significance, characteristics of the immunological milieu, and treatment responsiveness. Systematic screening with machine learning (ML) identified PSMB5 as a pivotal gene, with its functional importance further clarified using single-cell sequencing and Mendelian randomization analysis (MR). In vitro research confirmed the biological activities of PSMB5 in BLCA. Molecular docking demonstrated PSMB5's binding affinity with traditional Chinese medicines (TCMs). Clustering of 98 PANoptosis-associated genes revealed molecular subgroups A and B. A prognostic approach identified high-risk and low-risk cohorts, revealing considerable disparities in clinical characteristics and immunological landscapes across the groups. ML and MR identified PSMB5 as a risk factor in BLCA. Single-cell sequencing revealed that PSMB5 expression is predominantly associated with three cell lines linked to lymph node metastases. In vitro findings demonstrated that PSMB5 knockdown inhibited the proliferation and migration of BLCA cells while promoting apoptosis, whereas overexpression has the opposite effect. Molecular docking revealed a robust binding affinity between PSMB5 and five TCMs. A prognostic model incorporating PANoptosis-related genes was developed for stratifying BLCA risk and assessing the immune microenvironment. PSMB5 has been recognized as a crucial therapeutic target, exhibiting dual importance in the molecular etiology of BLCA and traditional Chinese medicine intervention.

Current risk stratification for stage II colorectal cancer (CRC) has limited accuracy in identifying patients who would benefit from adjuvant chemotherapy, leading to potential overtreatment or undertreatment. We aimed to develop a more precise risk stratification system by integrating artificial intelligence-based imaging analysis with pathological markers. We analyzed 2992 stage II CRC patients from 12 centers. A deep learning classifier (Swin Transformer Assisted Risk-stratification for CRC, STAR-CRC) was developed using multi-planar computed tomography (CT) images from 1587 patients (training : internal validation = 7 : 3) and validated in 1405 patients from eight independent centers, which stratified patients into low-, uncertain-, and high-risk groups. To further refine the uncertain-risk group, a composite score based on pathological markers (pT4 stage, number of lymph nodes sampled, perineural invasion, and lymphovascular invasion) was applied, forming the Intelligent Risk Integration System for stage II Colorectal Cancer (IRIS-CRC). IRIS-CRC was compared against the guideline-based risk stratification system (GRSS-CRC) for prediction performance and validated in the validation dataset. IRIS-CRC stratified patients into four prognostic groups with distinct 3-year disease-free survival rates (≥95%, 95%-75%, 75%-55%, ≤55%). Upon external validation, compared with GRSS-CRC, IRIS-CRC downstaged 27.1% of high-risk patients into the favorable group, and upstaged 6.5% of low-risk patients into the very poor prognosis group who might require more aggressive treatment. In the GRSS-CRC intermediate-risk group of the external validation dataset, IRIS-CRC reclassified 40.1% as favorable prognosis and 7.0% as very poor prognosis. IRIS-CRC's performance maintained generalizability in both chemotherapy and non-chemotherapy cohorts. IRIS-CRC offers a more precise and personalized risk assessment than current guideline-based risk factors, potentially sparing low-risk patients from unnecessary adjuvant chemotherapy while identifying high-risk individuals for more aggressive treatment. This novel approach holds promise for improving clinical decision-making and outcomes in stage II CRC.

Cancer is generally thought to be caused by expansion of a single mutant cell1. However, analyses of early colorectal cancer lesions indicate that tumours may instead originate from several genetically distinct cell populations2,3. Detecting polyclonal tumour initiation is challenging in patients, as it requires profiling early-stage lesions before clonal sweeps obscure diversity. To investigate this, we analysed normal colorectal mucosa, benign and dysplastic premalignant polyps and malignant adenocarcinomas (123 samples) from six individuals with familial adenomatous polyposis. Individuals with familial adenomatous polyposis have a germline heterozygous APC mutation, predisposing them to colorectal cancer and numerous premalignant polyps by early adulthood4. Whole-genome and/or whole-exome sequencing showed that many premalignant polyps-40% with benign histology and 28% with dysplasia-were composed of several genetic lineages that diverged early, consistent with polyclonal origins. This conclusion was reinforced by whole-genome sequencing of single crypts from polyps in further patients that showed limited sharing of mutations among crypts within the same lesion. In one case, several distinct APC mutations co-existed in different lineages of a single polyp, consistent with polyclonality. These findings reshape our understanding of early neoplastic events, demonstrating that tumour initiation can arise from the convergence of diverse mutant clones. They also indicate that cell-intrinsic growth advantages alone may not fully explain tumour initiation, highlighting the importance of microenvironmental and tissue-level factors in early cancer evolution.

Acute pancreatitis (AP) is a common disease, and several scores aim to assess its prognosis. Our study aims to automatically recognize mild AP from computed tomography (CT) images in patients with acute abdominal pain but uncertain diagnosis from clinical and serological data through Radiomic model based on formal methods (FMs). We retrospectively reviewed the CT scans acquired with Dual Source 256-slice CT scanner (Somatom Definition Flash; Siemens Healthineers, Erlangen, Germany) of 80 patients admitted to the radiology unit of Antonio Cardarelli hospital (Naples) with acute abdominal pain. Patients were divided into 2 groups: 40 underwent showed a healthy pancreatic gland, and 40 affected by four different grades (CTSI 0, 1, 2, 3) of mild pancreatitis at CT without clear clinical presentation or biochemical findings. Segmentation was manually performed. Radiologists identified 6 patients with a high expression of diseases (CTSI 3) to formulate a formal property (Rule) to detect AP in the testing set automatically. Once the rule was formulated, and Model Checker classified 70 patients into "healthy" or "unhealthy". The model achieved: accuracy 81%, precision 78% and recall 81%. Combining FMs results with radiologists agreement, and applying the mode in clinical practice, the global accuracy would have been 100%. Our model was reliable to automatically detect mild AP at primary diagnosis even in uncertain presentation and it will be tested prospectively in clinical practice.

Retrospectively compare image quality, radiologist diagnostic confidence, and time for images to reach PACS for contrast enhanced abdominopelvic CT examinations created on the scanner console by technologists versus those generated automatically by thin-client artificial intelligence (AI) mechanisms. A retrospective PACS search identified adults who underwent an emergency department contrast-enhanced abdominopelvic CT in 07/2022 (Console Cohort) and 07/2023 (Server Cohort). Coronal and sagittal multiplanar reformatted images (MPR) were created by AI software in the Server cohort. Time to completion of MPR images was compared using 2-sample t-tests for all patients in both cohorts. Two radiologists qualitatively assessed image quality and diagnostic confidence on 5-point Likert scales for 50 consecutive examinations from each cohort. Additionally, they assessed for acute abdominopelvic findings. Continuous variables and qualitative scores were compared with the Mann-Whitney U test. A p < .05 indicated statistical significance. Mean[SD] time to exam completion in PACS was 8.7[11.1] minutes in the Console cohort (n = 728) and 4.6[6.6] minutes in the Server cohort (n = 892), p < .001. 50 examinations in the Console Cohort (28 women 22 men, 51[19] years) and Server cohort (27 women 23 men, 57[19] years) were included for radiologist review. Age, sex, CTDlvol, and DLP were not statistically different between the cohorts (all p > .05). There was no significant difference in image quality or diagnostic confidence for either reader when comparing the Console and Server cohorts (all p > .05). Examinations utilizing AI generated MPRs on a thin-client architecture were completed approximately 50% faster than those utilizing reconstructions generated at the console with no statistical difference in diagnostic confidence or image quality.

Polyploid giant cancer cells (PGCCs) contribute to resistance against various cancer therapies. This study investigates whether HER2-directed antibody-drug conjugates (ADC) induce PGCCs and their role in drug resistance. HER2-positive breast cancer (JIMT-1) and gastric cancer (MKN7, SNU-216) cells were treated with HER2-directed ADCs, trastuzumab emtansine, trastuzumab deruxtecan, XMT-1522, and disitamab vedotin (DV). The induced persister cells were characterized using live cell imaging, confocal microscopy, immunohistochemistry, flow cytometry, gene expression analysis, SNP genotyping, and next-generation sequencing. Drug sensitivity was assessed using the AlamarBlue assay and SCID mouse xenografts. All 4 ADCs induced PGCCs, with XMT-1522 and DV being the most effective. The induced giant cells were drug-resistant and exhibited drug-tolerant persister cell characteristics. HER2 protein levels were downregulated in persisting drug-tolerant PGCCs and their daughter cells. JIMT-1 cells lost HER2 amplification following XMT-1522 treatment, along with the loss of extrachromosomal DNA containing HER2. However, XMT-1522-treated MKN7 and SNU-216 cells, and DV-treated JIMT-1 cells, retained the amplicon. Drug-tolerant PGCCs upregulated nectin-4, and treatment with enfortumab vedotin, a nectin-4-targeted ADC, inhibited the regrowth of JIMT-1 xenografts. ADC treatment induces PGCCs that contribute to drug resistance. ADC-induced drug-tolerant PGCCs express nectin-4, which may serve as a potential therapeutic target.

Objective: This article describes a new method (VS-FPM) for analysis of unstained tissues based on the application of supervised machine learning to generate brightfield hematoxylin and eosin (H&E) images from phase images recovered using Fourier ptychographic microscopy (FPM). Impact Statement: VS-FPM has several advantages for label-free digital pathology. Capture of complex image information simplifies model training and allows post-capture refocusing. FPM images combine high resolution with a large field of view, and the hardware is low-cost and compatible with many existing brightfield microscope systems. Introduction: By generating realistic histologically stained images from label-free image data, virtual staining (VS) methods have the potential to streamline clinical workflows, improve image consistency, and enable new ways of visualizing and analyzing histological tissues. Methods: We trained a conditional generative adversarial network to translate high-resolution FPM images of unstained tissues to brightfield H&E images and assessed the method using diagnosis of colonic polyps as a test case. Results: We found no statistically significant difference between the spatial resolution of FPM images captured at 4× magnification and images from a pathology slide scanner at 20× magnification. Visual assessment and image similarity metrics showed that VS-FPM images of unstained tissues closely resemble images of chemically H&E-stained tissues. However, the spatial resolution of virtual H&E images was approximately 20% lower than equivalent images of chemically stained tissues. Using VS-FPM, board-certified pathologists were able to accurately distinguish normal from dysplastic tissues and derive correct pathological diagnoses. Conclusion: VS-FPM is a reliable, accessible VS method that also overcomes many other limitations inherent to histopathology microscopy.

Purpose To develop a dentate nucleus (DN) segmentation tool using deep learning applied to brain MRI-based quantitative susceptibility mapping (QSM) images. Materials and Methods Brain QSM images from healthy controls and individuals with cerebellar ataxia or multiple sclerosis were collected from nine different datasets (2016-2023) worldwide for this retrospective study (ClinicalTrials.gov identifier: NCT04349514). Manual delineation of the DN was performed by experienced raters. Automated segmentation performance was evaluated against manual reference segmentations following training with several deep learning architectures. A two-step approach was used, consisting of a localization model followed by DN segmentation. Performance metrics included intraclass correlation coefficient (ICC), Dice score, and Pearson correlation coefficient. Results The training and testing datasets comprised 328 individuals (age range, 11-64 years; 171 female individuals), including 141 healthy individuals and 187 with cerebellar ataxia or multiple sclerosis. The manual tracing protocol produced reference standards with high intrarater (average ICC, 0.91) and interrater reliability (average ICC, 0.78). Initial deep learning architecture exploration indicated that the nnU-Net framework performed best. The two-step localization plus segmentation pipeline achieved a Dice score of 0.90 ± 0.03 (SD) and 0.89 ± 0.04 for left and right DN segmentation, respectively. In external testing, the proposed algorithm outperformed the current leading automated tool (mean Dice scores for left and right DN, 0.86 ± 0.04 vs 0.57 ± 0.22 [P < .001]; 0.84 ± 0.07 vs 0.58 ± 0.24 [P < .001]). The model demonstrated generalizability across datasets unseen during the training step, with automated segmentations showing high correlation with manual annotations (left DN: r = 0.74 [P < .001]; right DN: r = 0.48 [P = .03]). Conclusion The proposed model accurately and efficiently segmented the DN from brain QSM images. The model is publicly available (https://github.com/art2mri/DentateSeg). Keywords: MR Imaging, Brain/Brain Stem, Segmentation, Convolutional Neural Network, Supervised Learning, Computer Applications-3D, Volume Analysis, Image Postprocessing ClinicalTrials.gov registration no. NCT04349514 Supplemental material is available for this article. © RSNA, 2025.

Rationale: One in 10 children experiences sleep-disordered breathing (SDB). Untreated SDB is associated with poor cognition, but the underlying mechanisms are less understood. Objectives: We assessed the relationship between magnetic resonance imaging-derived upper airway volume and children's cognition and regional cortical gray matter volumes. Methods: We used 5-year data from the Adolescent Brain Cognitive Development study (N = 11,875 children; 9-10 yr old at baseline). Upper airway volumes were derived using a deep learning model applied to 5,552,640 brain magnetic resonance imaging slices. The primary outcome was the Total Cognition Composite score from the NIH Toolbox (NIH-TB). Secondary outcomes included other NIH-TB measures and cortical gray matter volumes. Measurements and Main Results: The habitual snoring group had significantly smaller airway volumes than nonsnorers (mean difference, 1.2 cm³; 95% confidence interval [CI], 1.0-1.4 cm³; P < 0.001). Deep learning-derived airway volume predicted the Total Cognition Composite score (estimated mean difference, 3.68 points; 95% CI, 2.41-4.96 points; P < 0.001) per one-unit increase in the natural log of airway volume (∼2.7-fold raw volume increase). This airway volume increase was also associated with an average 0.02-cm³ increase in right temporal pole volume (95% CI, 0.01-0.02 cm³; P < 0.001). Similar airway volume predicted most NIH-TB domain scores and multiple frontal and temporal gray matter volumes. These brain volumes mediated the relationship between airway volume and cognition. Conclusions: We demonstrate a novel application of deep learning-based airway segmentation in a large pediatric cohort. Upper airway volume is a potential biomarker for cognitive outcomes in pediatric SDB, offers insights into neurobiological mechanisms, and informs future studies on risk stratification.

The aim of the study was to compare the diagnostic quality of deep learning (DL) reconstructed balanced steady-state free precession (bSSFP) single-shot (SSH) cine images with standard, multishot (also: segmented) bSSFP cine (standard cine) in cardiac MRI. This prospective study was performed in a cohort of participants with clinical indication for cardiac MRI. SSH compressed-sensing bSSFP cine and standard multishot cine were acquired with breath-holding and electrocardiogram-gating in short-axis view at 1.5 Tesla. SSH cine images were reconstructed using an industry-developed DL super-resolution algorithm (DL-SSH cine). Two readers evaluated diagnostic quality (endocardial edge definition, blood pool to myocardium contrast and artifact burden) from 1 (nondiagnostic) to 5 (excellent). Functional left ventricular (LV) parameters were assessed in both sequences. Edge rise distance, apparent signal-to-noise ratio (aSNR) and contrast-to-noise ratio were calculated. Statistical analysis for the comparison of DL-SSH cine and standard cine included the Student's t-test, Wilcoxon signed-rank test, Bland-Altman analysis, and Pearson correlation. Forty-five participants (mean age: 50 years ±18; 30 men) were included. Mean total scan time was 65% lower for DL-SSH cine compared to standard cine (92 ± 8 s vs 265 ± 33 s; P  < 0.0001). DL-SSH cine showed high ratings for subjective image quality (eg, contrast: 5 [interquartile range {IQR}, 5-5] vs 5 [IQR, 5-5], P  = 0.01; artifacts: 4.5 [IQR, 4-5] vs 5 [IQR, 4-5], P  = 0.26), with superior values for sharpness parameters (endocardial edge definition: 5 [IQR, 5-5] vs 5 [IQR, 4-5], P  < 0.0001; edge rise distance: 1.9 [IQR, 1.8-2.3] vs 2.5 [IQR, 2.3-2.6], P  < 0.0001) compared to standard cine. No significant differences were found in the comparison of objective metrics between DL-SSH and standard cine (eg, aSNR: 49 [IQR, 38.5-70] vs 52 [IQR, 38-66.5], P  = 0.74). Strong correlation was found between DL-SSH cine and standard cine for the assessment of functional LV parameters (eg, ejection fraction: r = 0.95). Subgroup analysis of participants with arrhythmia or unreliable breath-holding (n = 14/45, 31%) showed better image quality ratings for DL-SSH cine compared to standard cine (eg, artifacts: 4 [IQR, 4-5] vs 4 [IQR, 3-5], P  = 0.04). DL reconstruction of SSH cine sequence in cardiac MRI enabled accelerated acquisition times and noninferior diagnostic quality compared to standard cine imaging, with even superior diagnostic quality in participants with arrhythmia or unreliable breath-holding.

This study aims to assess diagnostic performance of high-resolution dynamic contrast-enhanced (DCE) MRI with deep learning-based compressed sensing and super-resolution (DLCS-SR) reconstruction for identifying microadenomas. This prospective study included 126 participants with suspected pituitary microadenomas who underwent DCE MRI between June 2023 and January 2024. Four image groups were derived from single-scan DCE MRI, which included 1.5-mm slice thickness images using DLCS-SR (1.5-mm DLCS-SR images), 1.5-mm slice thickness images with deep learning-based compressed sensing reconstruction (1.5-mm DLCS images), 1.5-mm routine images, and 3-mm slice thickness images using DLCS-SR (3-mm DLCS-SR images). Diagnostic criteria were established by incorporating laboratory findings, clinical symptoms, medical histories, previous imaging, and certain pathologic reports. Two readers assessed the diagnostic performance in identifying pituitary abnormalities and microadenomas. Diagnostic agreements were assessed using κ statistics, and intergroup comparisons for microadenoma detection were performed using the DeLong and McNemar tests. The 1.5-mm DLCS-SR images (κ = 0.746-0.848) exhibited superior diagnostic agreement, outperforming 1.5-mm DLCS (κ = 0.585-0.687), 1.5-mm routine (κ = 0.449-0.487), and 3-mm DLCS-SR images (κ = 0.347-0.369) (p < 0.001 for all). Additionally, the performance of 1.5-mm DLCS-SR images in identifying microadenomas [area under the receiver operating characteristic curve (AUC), 0.89-0.94] surpassed that of 1.5-mm DLCS (AUC, 0.83-0.87; p = 0.042 and 0.011, respectively), 1.5-mm routine (AUC, 0.76-0.78; p < 0.001), and 3-mm DLCS-SR images (AUC, 0.72-0.74; p < 0.001). The findings revealed superior diagnostic performance of 1.5-mm DLCS-SR images in identifying pituitary abnormalities and microadenomas, indicating the clinical-potential of high-resolution DCE MRI. Question What strategies can overcome the resolution limitations of conventional dynamic contrast-enhanced (DCE) MRI, and which contribute to a high false-negative rate in diagnosing pituitary microadenomas? Findings Deep learning-based compressed sensing and super-resolution reconstruction applied to DCE MRI achieved high resolution while improving image quality and diagnostic efficacy. Clinical relevance DCE MRI with a 1.5-mm slice thickness and high in-plane resolution, utilizing deep learning-based compressed sensing and super-resolution reconstruction, significantly enhances diagnostic accuracy for pituitary abnormalities and microadenomas, enabling timely and effective patient management.

Electron microscopy is an important technique for the study of synaptic morphology and its relation to synaptic function. The data analysis for this task requires the segmentation of the relevant synaptic structures, such as synaptic vesicles (SV), active zones, mitochondria, presynaptic densities, synaptic ribbons, and synaptic compartments. Previous studies were predominantly based on manual segmentation, which is very time-consuming and prevented the systematic analysis of large datasets. Here, we introduce SynapseNet, a tool for the automatic segmentation and analysis of synapses in electron micrographs. It can reliably segment SVs and other synaptic structures in a wide range of electron microscopy approaches, thanks to a large annotated dataset, which we assembled, and domain adaptation functionality we developed. We demonstrated its capability for (semi-)automatic biological analysis in two applications and made it available as an easy-to-use tool to enable novel data-driven insights into synapse organization and function.

Purpose To assess the effectiveness of an explainable deep learning model, developed using multiparametric MRI features, in improving diagnostic accuracy and efficiency of radiologists for classification of focal liver lesions (FLLs). Materials and Methods FLLs 1 cm or larger in diameter at multiparametric MRI were included in the study. The nn-Unet and Liver Imaging Feature Transformer models were developed using retrospective data from the Ruijin Hospital (January 2018-August 2023). The nnU-Net was used for lesion segmentation and the Liver Imaging Feature Transformer model for FLL classification. External testing was performed on data from the Xinjiang Production and Construction Corps Hospital, the First Affiliated Hospital of Soochow University, and Xinrui Hospital (January 2018-December 2023), with a prospective test set obtained from January to April 2024. Model performance was compared with radiologists, and impact of model assistance on junior and senior radiologist performance was assessed. Evaluation metrics included the Dice similarity coefficient and accuracy. Results A total of 2131 individuals with FLLs (mean age, 56 years ± 12 [SD]; 1476 female patients) were included in the training, internal test, external test, and prospective test sets. Average Dice similarity coefficient values for liver and tumor segmentation across the three test sets were 0.98 and 0.96, respectively. Average accuracy for features and lesion classification across the three test sets were 93% and 97%, respectively. Readings assisted by the Liver Imaging Feature Transformer model improved diagnostic accuracy (average 5.3% increase, P < .001), reduced reading time (average 34.5 seconds decrease, P < .001), and enhanced confidence (average 0.3-point increase, P < .001) of junior radiologists. Conclusion The proposed deep learning model accurately detected and classified FLLs, improving diagnostic accuracy and efficiency of junior radiologists. Keywords: Liver, MR-Dynamic Contrast Enhanced, Convolutional Neural Network (CNN), Deep Learning Algorithms, Machine Learning Algorithms, Feature Detection, Vision, Application Domain Supplemental material is available for this article. © RSNA, 2025 See also commentary by Adams and Bressem in this issue.

A heart-convolutional neural network (heart-CNN) was developed and tested for the automatic detection of left atrial enlargement (LAE) from feline thoracic radiographs. A retrospective and multicenter study was performed. Right lateral and dorso-ventral and/or ventro-dorsal thoracic radiographs of cats with concomitant echocardiographic examination were selected from the internal databases of both academic and private referral institutions. Radiographic images were classified as no LAE, mild, moderate and severe LAE, based on echocardiographic reports. Heart-CNN performance was evaluated using confusion matrices and receiver operating characteristic curves for both radiographic projections considering a multiclass and a binary classification. Considering the multiclass classification, for the right lateral view, the area under the curve (AUC) was of 0.73, 0.68, 0.64 and 0.78 for the no LAE, mild, moderate and severe LAE groups, respectively. The AUCs for the dorso-ventral and/or ventro-dorsal images were 0.73, 0.64, 0.63 and 0.76 for the no LAE, mild, moderate and severe LAE groups, respectively. In the binary classification, AUCs were 0.83 and 0.81 for right lateral and dorso-ventral and/or ventro-dorsal projections, respectively. The developed AI-based tool seems to be a promising support for automatic identification of more advanced stages of LAE in cats.

This systematic review assesses the diagnostic accuracy, feasibility, and clinical performance of artificial intelligence (AI)-based smartphone imaging tools for detecting dental caries. Following the Preferred Reporting Items for Systematic Reviews and Meta-Analyses for Diagnostic Test Accuracy (PRISMA-DTA) guidelines, five databases: PubMed, Scopus, Web of Science, Embase, and Cochrane Library, were searched up to March 26, 2025. This study was registered with the International Prospective Register of Systematic Reviews (PROSPERO) (CRD420251047689). Diagnostic accuracy and feasibility of AI-driven analysis of smartphone-based dental images for the detection of dental caries were assessed. Risk of bias and applicability were evaluated using QUADAS-2. Fourteen studies met the inclusion criteria. AI models, particularly YOLO variants, DenseNet201, and MobileNetV3, demonstrated high diagnostic accuracy, especially for cavitated lesions, with some outperforming junior dentists. Enhanced YOLO models achieved up to 85.5 % mean average precision. Tools were generally user-friendly and suitable for community or at-home screening. However, sensitivity for early or non-cavitated lesions varied. AI-driven smartphone imaging shows promise as an accessible and reliable tool for caries detection, particularly in low-resource or remote settings. Further research is needed to improve early lesion detection, ensure clinical validation, and support equitable implementation.

Diagnosis and surveillance of bladder cancer rely on white-light cystoscopy (WLC). However, this modality is operator-dependent and associated with a risk of missed lesions, contributing to high recurrence rates, especially in non-muscle invasive bladder cancer. Recent advances in artificial intelligence (AI) enable software-based decision support for bladder lesion detection, with potential for vendor-independent deployment and broad integration into routine clinical workflows. To develop and externally validate an AI-based clinical decision support system for real-time bladder lesion detection during cystoscopy. CystoAID, a convolutional neural network-based object detection system, was trained on prospectively collected video recordings from flexible cystoscopies and transurethral resections of bladder tumors. Diagnostic accuracy was evaluated using a retrospective external validation dataset representative of routine clinical practice, in accordance with STARD-AI recommendations. In the external validation cohort, CystoAID achieved a sensitivity of 1.00 (95% CI 0.95-1.00). Precision was 88.1% (95% CI 81.3-92.7), exceeding published estimates for WLC. Precision-recall analysis showed consistently high precision (>0.8) across clinically relevant recall levels, with declining precision at higher recall, reflecting the expected trade-off between sensitivity and false-positive detections. The system operated with low processing latency, supporting feasibility for real-time clinical use. Sensitivity was prioritized to mitigate the clinical risk associated with false-negative findings. CystoAID is a real-time, AI-based decision support tool for cystoscopy that demonstrated high sensitivity and favorable precision in external validation. These findings support its potential role as an assistive technology in routine urologic practice. Prospective studies are warranted to evaluate clinical impact, workflow integration, and performance in detecting challenging lesion subtypes, including flat lesions and carcinoma in situ.

To develop an automated analytical framework that integrates hybrid radiomics and deep learning features from non-contrast CT images for the multi-class classification of bladder pathologies. This retrospective study analyzed 902 CT scans (584 normal, 142 calculi, 66 cancers, 110 cystitis). An integrated pipeline was implemented, comprising: 1) automatic bladder segmentation using a 3D-UNet, 2) hybrid feature extraction combining 100 radiomics features and 256 deep features from a 3D convolutional autoencoder, 3) feature selection via variance thresholding and LASSO regression, and 4) final classification using an XGBoost classifier. The dataset was split into training (80 %) and validation (20 %) sets. Performance was evaluated using the area under the receiver operating characteristic curve (AUROC) with a one-vs-rest strategy for multi-class classification. Model stability was assessed via stratified five-fold cross-validation, and interpretability was analyzed with SHapley Additive exPlanations (SHAP). The framework achieved one-vs-rest AUROCs of 0.94 (95 % CI: 0.89-0.99) for calculi, 0.92 (0.85-0.99) for cancer, 0.90 (0.84-0.95) for normal bladder, and 0.83 (0.75-0.91) for cystitis. The micro-average AUROC for four-class discrimination was 0.94 (0.92-0.96). Binary normal/abnormal classification demonstrated stable performance across cross-validation folds (AUROC range: 0.89-0.92). SHAP analysis revealed that radiomic features dominated decisions for calculi/normal differentiation, while deep features were critical for distinguishing cancer and cystitis. The proposed hybrid CT analysis framework achieves clinically relevant performance in the automated, multi-class classification of bladder pathologies, excelling particularly in calculi detection. The complementary roles of radiomic and deep features provide an interpretable diagnostic aid, demonstrating potential for integration into clinical workflows to support differential diagnosis.

This study develops and validates a deployable, privacy-preserving automated tool to address the labor-intensive nature of colorectal cancer (CRC) recurrence registration. Our tool utilizes a reproducible, two-stage 'Clinical Context Engineering' workflow to mimic expert clinical reasoning and overcome the limitations of handling longitudinal clinical data ambiguity. We retrospectively studied 3053 CRC patients (2010-2018). A local Large Language Model (LLM) (Qwen3:14B) analyzed ∼ 19,900 pathology and ∼ 43,900 imaging reports using iterative, patch-based analysis ("raw LLM"). Clinical validation rules were applied to generate "rule-based LLM" output, enhancing explainability and trustworthiness. Both automated methods and the manual Taiwan Cancer Registry (TCR) database were compared against a 20% manual reference standard (N = 602). Full prompts and validation code are provided for complete reproducibility. Under a strict 60-day temporal tolerance, the Rule-Based LLM achieved 90.7% accuracy, comparable to standard TCR processes (92.0%), and 77.2% sensitivity. The application of clinical validation rules significantly improved specificity from 87.7% (Raw LLM) to 93.9% (Rule-Based LLM). In time window analysis, the Rule-Based LLM identified 87.1% of recurrences within 60 days of the reference date. Our locally deployed, privacy-preserving, and explainable Clinical Context Engineering framework offers a viable, non-inferior alternative to standard TCR processes, reducing workload while maintaining data quality and fostering trust in AI-assisted cancer registry automation.

Accurate preoperative staging of colorectal cancer is critical to guide treatment decisions, including eligibility for R0 resection, to reduce recurrence and improve patient survival. However, conventional imaging evaluation depends on human experience, leading to subjectivity and diagnostic uncertainty in clinical interpretation. Therefore, an objective, evidence-based quantitative analysis tool is essential to assist clinicians. This study introduces and validates an evidence-based medicine (EBM) deep learning model that addresses these limitations by providing reliable, automated pathological screening from medical images. We formulate the novel EASDnet model, an EBM-based deep learning architecture for the discriminative screening of colorectal cancer lesions, which incorporates an evidence and attention-based mechanism to learn subtle morphological differences between cancerous lesions and their surrounding microenvironment. This approach effectively captures inter-class and intra-class discriminative and differentiating features from histopathological data. EASDnet is rigorously trained and validated using publicly available image datasets, NCT-100K and LC25000. The proposed model demonstrated robust performance in quantitative colorectal cancer diagnosis. Evaluating EASDnet on the respective datasets confirmed its high discriminative capability, with accuracy scores of 97.76% and 98.52%. The pathological diagnostic capabilities of EASDnet are superior to the current state-of-the-art methods for determining colorectal cancer status. With its optimized performance in tissue dividing and feature inference, EASDnet is a promising and valuable learning model that can advance objective diagnostic strategies and improve clinical care for patients with colorectal cancer.

Abdominal radiography (AXR) is routinely performed in emergency departments (ED) but has limited clinical utility. Thus, this study developed deep-learning models using AXR to predict hospitalization in ED patients with abdominal symptoms. This retrospective study included 1,585 adult ED patients with abdominal symptoms who underwent AXR between August and December 2021. External validation included 112 patients. Three prediction models were developed using random forest classifiers: an image model, a clinical model, and a fusion model combining both image and clinical features. DenseNet201 extracted image features, and early clinical information obtainable by non-physician staff was incorporated. The performances of the two radiologists were compared with those of a deep-learning model. The fusion model achieved an area under the receiver operating characteristic curve (AUROC) of 0.70 (95% confidence interval: 0.65-0.76), sensitivity of 0.82, and an F1-score of 0.75 in internal validation. Human readers showed high specificity (0.80-0.95) but low sensitivity (0.13-0.52) and F1-scores (0.23-0.55). Deep-learning models achieved substantially higher sensitivity and F1-scores than human readers, but with lower specificity (0.43), making them suitable for screening. In the external validation, the performance of the fusion model decreased (AUROC, 0.60), but image-based models maintained a higher sensitivity than human readers. Of all the radiographs, 11.1% showed abnormal findings, and 5.8% were specific to the final diagnosis. Deep learning significantly improved the clinical utility of AXR in screening for hospitalization risk, achieving higher sensitivity and F1-scores than radiologists. While these models require further improvement for clinical implementation, they can potentially extract predictive patterns from traditionally limited imaging studies.

Artificial intelligence (AI) is increasingly used in radiology, but its development in pediatric imaging remains limited, particularly for emergent conditions. Ileocolic intussusception is an important cause of acute abdominal pain in infants and toddlers and requires timely diagnosis to prevent complications such as bowel ischemia or perforation. While ultrasonography is the diagnostic standard due to its high sensitivity and specificity, its accessibility may be limited, especially outside tertiary centers. Abdominal radiographs (AXRs), despite their limited sensitivity, are often the first-line imaging modality in clinical practice. In this context, AI could support early screening and triage by analyzing AXRs and identifying patients who require further ultrasonography evaluation. This study aimed to upgrade and externally validate an AI model for screening ileocolic intussusception using pediatric AXRs with multicenter data and to assess the diagnostic performance of the model in comparison with radiologists of varying experience levels with and without AI assistance. This retrospective study included pediatric patients (≤5 years) who underwent both AXRs and ultrasonography for suspected intussusception. Based on the preliminary study from hospital A, the AI model was retrained using data from hospital B and validated with external datasets from hospitals C and D. Diagnostic performance of the upgraded AI model was evaluated using sensitivity, specificity, and the area under the receiver operating characteristic curve (AUC). A reader study was conducted with 3 radiologists, including 2 trainees and 1 pediatric radiologist, to evaluate diagnostic performance with and without AI assistance. Based on the previously developed AI model trained on 746 patients from hospital A, an additional 431 patients from hospital B (including 143 intussusception cases) were used for further training to develop an upgraded AI model. External validation was conducted using data from hospital C (n=68; 19 intussusception cases) and hospital D (n=90; 30 intussusception cases). The upgraded AI model achieved a sensitivity of 81.7% (95% CI 68.6%-90%) and a specificity of 81.7% (95% CI 73.3%-87.8%), with an AUC of 86.2% (95% CI 79.2%-92.1%) in the external validation set. Without AI assistance, radiologists showed lower performance (overall AUC 64%; sensitivity 49.7%; specificity 77.1%). With AI assistance, radiologists' specificity improved to 93% (difference +15.9%; P<.001), and AUC increased to 79.2% (difference +15.2%; P=.05). The least experienced reader showed the largest improvement in specificity (+37.6%; P<.001) and AUC (+14.7%; P=.08). The upgraded AI model improved diagnostic performance for screening ileocolic intussusception on pediatric AXRs. It effectively enhanced the specificity and overall accuracy of radiologists, particularly those with less experience in pediatric radiology. A user-friendly software platform was introduced to support broader clinical validation and underscores the potential of AI as a screening and triage tool in pediatric emergency settings.

Oral cavity squamous cell carcinoma (OSCC) is a global health burden, where negative margins are essential for reducing recurrence and improving survival. Intraoperative frozen-section analysis is limited by time, sampling error, and interpretive variability, underscoring the need for more reliable margin assessment. Reflectance confocal microscopy (RCM) enables real-time, in vivo high-resolution imaging, but accuracy depends on expert interpretation. This study evaluated the diagnostic performance of an artificial intelligence (AI)-driven model for RCM in OSCC, aiming to develop a point-of-care platform for intraoperative use. Patients with biopsy-confirmed OSCC underwent in vivo RCM imaging using a handheld intraoral probe before biopsy. Histopathology was the reference standard. A deep learning model was developed with the Google Cloud Vertex AI Automated Machine Learning (AutoML) Vision platform and trained on 4,090 annotated RCM images (1,998 benign, 2,092 malignant). Performance was compared with blinded expert pathologist and RCM readers. The AI model achieved an area under the precision-recall curve (AUC-PR) of 0.99 and an area under the receiver operating characteristic curve (AUC-ROC) of 0.99, with sensitivity 98.09%, specificity 95.00%, accuracy 96.58%, positive predictive value (PPV) 95.35%, and negative predictive value (NPV) 97.94%. Expert readers showed sensitivity 90.00%, specificity 98.30%, accuracy 94.15%, PPV 88.20%, and NPV 96.60%. Inter-reader agreement was 95.00% for benign and 81.70% for malignant cases. AI-driven RCM interpretation provides an accurate, rapid, noninvasive approach for OSCC diagnosis and intraoperative margin assessment. It outperformed expert readers and can reduce reliance on frozen-section analysis, streamline workflows, and improve outcomes.

The objective of this study was to investigate the antibacterial and demineralisation-inhibiting effects of a non-discolouring silver complex fluoride (SCF) on artificial dentine caries. Dentine blocks with artificial carious lesions were treated with SCF, silver diamine fluoride (SDF, as positive control), and water (as negative control). Then the blocks underwent Streptococcus mutans biofilm and pH-cycling for 7 days. The morphology, viability and growth kinetics of the biofilm were evaluated by scanning electron microscopy (SEM), confocal laser scanning microscopy (CLSM) and colony-forming unit (CFU) counting, respectively. The dentine surface morphology, lesion depths and mineral loss, chemical structure and crystal characteristics were determined using SEM, micro-computed tomography, Fourier transform infrared (FTIR), and X-ray diffraction (XRD), respectively. One-way analysis of variance with the Bonferroni post hoc test was performed to assess and compare the data. SEM revealed confluent bacterial growth covering the surface of dentine treated with water but not SCF and SDF. CLSM showed that the dead-to-live ratios of biofilms treated with SCF, SDF, and water were 0.77 ± 0.16, 0.88 ± 0.40, and 0.44 ± 0.03, respectively (p < 0.001, SCF, SDF>Water). The Log CFU values of the biofilm treated with SCF, SDF and water were 6.7 ± 0.1, 6.6 ± 0.2 and 7.8 ± 0.1 (p < 0.001, SCF, SDF<Water). SEM images showed Group SCF had less exposed dentine collagen fibers than Group Water. Micro-computed tomography showed the lesion-depth (μm) of dentine treated with SCF, SDF and water were 40±5, 35±8 and 180±20 (p < 0.001, SCF, SDF<Water). The mineral loss (gHApcm-3) of dentine treated with SCF, SDF and water were 0.24±0.05, 0.20±0.03 and 0.45±0.03 (p < 0.001, SCF, SDF<Water). FTIR showed the amide I-to-hydrogen phosphate ratios of dentine treated with SCF, SDF and water were 0.25±0.03, 0.22±0.03, and 0.43±0.05 (p < 0.001, SCF, SDF<Water). XRD revealed that the hydroxyapatite in SDF and SCF groups was better crystallised than that of water Group. This study demonstrated that SCF inhibited the growth of S. mutans biofilm and reduced the demineralisation of artificial dentine caries. If SCF is successfully translated into clinical application, it may be a novel anti-caries agent for clinicians to arrest dentine caries.

Implantable neurotechnologies are increasingly used to reduce seizure burden in pediatric epilepsy. Vagus nerve stimulation (VNS), the most common option, is effective for only half of patients, with no means to predict outcome prior to surgery. As a result, many children undergo invasive and costly procedures without benefit. Although T1-weighted magnetic resonance imaging (T1w) is routinely acquired presurgically and may capture structural brain differences relevant to treatment outcome, its high dimensionality relative to sample sizes has limited its utility in predictive modelling. To address this challenge, we present VQ-VNS, a deep representation learning model to predict VNS outcome based on preoperative T1w (n = 263). First, we present data from the largest paediatric VNS cohort (n = 1046), wherein presurgical clinical data could not predict response (AUC 0.54,p > 0.99). Next, VQ-VNS was pretrained on 7433 T1w images to learn compact anatomical representations enabling its classifier to predict VNS response (AUC = 0.73,p = 0.007). Model predictions localized to serotonin-rich brain regions and inferred large-scale disruptions in network connectivity among non-responders. This biologically interpretable predictor based on routine structural imaging improves upon current clinical decision-making.

Motion and limited compliance compromise diagnostic MR image quality, particularly in pediatric patients who frequently require sedation. Single-shot sequences offer a time-efficient alternative but suffer from reduced image quality. This study aimed to evaluate the diagnostic performance of a deep learning (DL) framework combining compressed sensing (CS) and convolutional neural networks (CNNs) to enhance T2-weighted single-shot MRI (T2-SSHDL) compared with conventional CS-based reconstruction (T2-SSHconv) and routinely acquired high-resolution T2-weighted sequences. This prospective single-center study included 62 pediatric patients (mean age, 7.4 ± 4.9 years; 36 males, 26 females), who underwent T2-weighted single-shot brain MRI (29 sedated, 33 awake). Raw data were reconstructed using a DL-based pipeline and compared with conventional CS-based reconstructions. Quantitative metrics included apparent contrast-to-noise ratio (aCNR), apparent signal-to-noise ratio (aSNR), and edge rise distance (ERD). Two radiologists rated images for artifacts, sharpness, lesion conspicuity, and overall quality on a 5-point Likert scale. T2-SSHDL-sequences showed significantly higher aCNR (29.9 ± 22.6 vs. 26.7 ± 16.5; p < 0.001), aSNR (41.6 ± 27.9 vs. 38.2 ± 20.8; p = 0.003), and improved sharpness (ERD 0.90 ± 0.35 mm vs. 1.35 ± 0.42 mm; p < 0.001). Qualitative assessments confirmed superior image quality, lesion conspicuity, and sharpness (p < 0.001). Compared with high-resolution T2-weighted sequences, T2-SSHDL-sequences showed fewer motion artifacts and comparable lesion conspicuity in non-sedated patients. DL-based reconstruction significantly enhances the diagnostic quality of T2-weighted single-shot brain MRI in pediatric patients, enabling clinically usable, ultrafast, motion-robust imaging with potential to reduce the need for sedation. Question Can deep learning-based reconstruction elevate motion-robust single-shot T2-weighted pediatric brain MRI to diagnostic image quality levels, enabling reliable imaging without sedation? Findings Both quantitative and qualitative evaluations confirmed significantly improved image quality of deep learning-enhanced single-shot T2-weighted brain MRI compared with conventional reconstruction. Clinical relevance Deep learning-enhanced reconstruction improves image quality in ultrafast, motion-robust single-shot pediatric brain MRI, potentially reducing the need for sedation while preserving diagnostic accuracy. This approach may enhance patient safety and shorten examination time in routine neuroimaging.

Accurate identification of brain metastases is critical for determining prognosis and guiding treatment. Deep learning reconstruction (DLR) enhances MRI quality by reducing noise, while super-resolution DLR (SR-DLR) may further improve spatial resolution and lesion detectability. To evaluate SR-DLR versus conventional DLR in detecting and visualizing brain metastases on postcontrast T1-weighted brain MRI. This retrospective study included 47 consecutive patients who underwent postcontrast 3D whole-brain T1-weighted MRI between July and December 2024. Images were reconstructed using both SR-DLR and DLR. Three independent readers evaluated metastatic lesion detection and rated image quality. Subjective assessments included lesion visibility, visibility of normal structures, sharpness, noise, and overall image quality. Objective metrics-full width at half maximum (FWHM), edge rise distance (ERD), edge rise slope (ERS), signal-to-noise ratio (SNR), and contrast-to-noise ratio (CNR)-were also measured. Statistical tests included jackknife alternative free-response receiver operating characteristic (JAFROC) analysis, Wilcoxon signed-rank test, McNemar's test, and paired t-tests, with significance threshold of p < 0.050. A total of 117 brain metastases were detected in 47 patients (mean age, 59 years ± 18; 27 men). SR-DLR demonstrated significantly better lesion detection performance than DLR (mean figure of merit = 0.842 vs. 0.797; p = 0.042). Subjective image quality ratings favored SR-DLR for lesion and structure visibility, sharpness, noise, and overall quality in most cases. Objectively, SR-DLR yielded lower FWHM (1.2 mm vs. 1.9 mm; p < 0.001), higher ERS (791.3 mm- 1 vs. 645.3 mm- 1; p = 0.013) indicating enhanced sharpness as well as improved CNR (27.5 vs. 24.9; p < 0.001) compared to DLR. Compared to DLR, SR-DLR significantly enhances brain MRI quality and improves detection of metastatic lesions.

Diffusion magnetic resonance imaging (dMRI) often suffers from low spatial and angular resolution due to inherent limitations in imaging hardware and system noise, adversely affecting the accurate estimation of microstructural parameters with fine anatomical details. Deep learning-based super-resolution techniques have shown promise in enhancing dMRI resolution without increasing acquisition time. However, most existing methods are confined to either spatial or angular super-resolution, disrupting the information exchange between the two domains and limiting their effectiveness in capturing detailed microstructural features. Furthermore, traditional pixel-wise loss functions only consider pixel differences, and struggle to recover intricate image details essential for high-resolution reconstruction. We propose SHRL-dMRI, a novel Spherical Harmonics Representation Learning framework for high-fidelity, generalizable super-resolution in dMRI to address these challenges. SHRL-dMRI explores implicit neural representations and spherical harmonics to model continuous spatial and angular representations, simultaneously enhancing both spatial and angular resolution while improving the accuracy of microstructural parameter estimation. To further preserve image fidelity, a data-fidelity module and wavelet-based frequency loss are introduced, ensuring the super- resolved images preserve image consistency and retain fine details. Extensive experiments demonstrate that, compared to five other state-of-the-art methods, our method significantly enhances dMRI data resolution, improves the accuracy of microstructural parameter estimation, and provides better generalization capabilities. It maintains stable performance even under a 45× downsampling factor. The proposed method can effectively improve the resolution of dMRI data without increasing the acquisition time, providing new possibilities for future clinical applications.

Reinforcement learning (RL) has emerged as a powerful artificial intelligence paradigm in medical image analysis, excelling in complex decision-making tasks. This systematic review synthesizes the applications of RL across diverse imaging domains-including landmark detection, image segmentation, lesion identification, disease diagnosis, and image registration-by analyzing 20 peer-reviewed studies published between 2019 and 2023. RL methods are categorized into classical and deep reinforcement learning (DRL) approaches, focusing on their performance, integration with other machine learning models, and clinical utility. Deep Q-Networks (DQN) demonstrated strong performance in anatomical landmark detection and cardiovascular risk estimation, while Proximal Policy Optimization (PPO) and Advantage Actor-Critic (A2C) achieved optimal policy learning for vessel tracking. Policy gradient methods such as REINFORCE, Twin-Delayed Deep Deterministic Policy Gradient (TD3), and Soft Actor-Critic (SAC) were successfully applied to breast lesion detection, white-matter connectivity analysis, and vertebral segmentation.Monte Carlo learning, meta-RL, and A3C methods proved effective for adaptive questioning, image quality evaluation, and multimodal image registration. To consolidate these findings, we propose a unified Reinforcement Learning Medical Imaging (RLMI) framework encompassing four core components: state representation, policy optimization, reward formulation, and environment modeling. This framework enhances sequential agent learning, stabilizes navigation, and generalizes across imaging modalities and tasks. Key challenges remain, including optimizing task-specific policies, integrating anatomical contexts, addressing data scarcity, and improving interpretability. This review highlights RL's potential to enhance accuracy, adaptability, and efficiency in medical image analysis, providing valuable guidance for researchers and clinicians applying RL in real-world healthcare settings.

Large vision-language models (VLMs), such as GPT-4.1 and GPT-4o, have shown strong potential in medical tasks. However, their application in lossless 3D medical image analysis is still underexplored. This study proposes a GPT-based diagnostic approach that maintains voxel-level accuracy during data ingestion, structures multi-slice visual inputs for model interpretation, and integrates consensus guidelines to align predictions with clinical standards. In doing so, the approach may provide interpretable and guideline-consistent decision support even for less experienced clinicians. We designed a novel approach that directly processes 3D chest CT scans in NIfTI format, maintains full voxel fidelity during data import and analysis, and is compatible with GPT-based workflows. For each lung nodule, we guided GPT in analyzing multi-slice visual inputs, including bounding annotations, segmentation overlays, and cropped views. Guidelines (Fleischner, BTS, ACCP) were embedded to promote standardized interpretation and guide reasoning from nodule-level characteristics to patient-level assessment. Three NIfTI-based input settings were used to test the method on the LIDC-IDRI dataset: (1) nodule coordinates only; (2) coordinates with guideline-based prompting; (3) segmentation overlays with guideline-based prompting. To evaluate the performance on external datasets, we performed external validation on the Lung Nodule Database (LNDb). GPT-4.1 outperformed GPT-4o overall, especially with full-context input, while GPT-4o had higher sensitivity with limited input. With segmentation and guideline-based prompting, GPT-4.1 achieved accuracy 0.722 and AUC 0.780 on LIDC-IDRI dataset. In external validation on LNDb dataset, it reached accuracy 0.767 and AUC 0.780. GPT-4.1 maintained its competitiveness when compared to representative deep-learning baselines and radiologist readers. It also provided stronger interpretability through guideline-grounded and patient-level reasoning with explicit textual justifications. This study presents a clinically aligned and interpretable approach for GPT-based lung cancer diagnosis using lossless 3D CT images. The outcomes demonstrate the potential of combining large vision-language models with structured visual and guideline-based context in real-world diagnostic workflows. The online version contains supplementary material available at 10.1007/s13755-025-00417-8.

The rising incidence of pancreatic diseases, including acute and chronic pancreatitis and various pancreatic neoplasms, poses a significant global health challenge. Pancreatic ductal adenocarcinoma (PDAC) for example, has a high mortality rate due to late-stage diagnosis and its inaccessible location. Advances in imaging technologies, though improving diagnostic capabilities, still necessitate biopsy confirmation. Artificial intelligence, particularly machine learning and deep learning, has emerged as a revolutionary force in healthcare, enhancing diagnostic precision and personalizing treatment. This narrative review explores Artificial intelligence's role in pancreatic imaging, its technological advancements, clinical applications, and associated challenges. Following the PRISMA-DTA guidelines, a comprehensive search of databases including PubMed, Scopus, and Cochrane Library was conducted, focusing on Artificial intelligence, machine learning, deep learning, and radiomics in pancreatic imaging. Articles involving human subjects, written in English, and published up to March 31, 2024, were included. The review process involved title and abstract screening, followed by full-text review and refinement based on relevance and novelty. Recent Artificial intelligence advancements have shown promise in detecting and diagnosing pancreatic diseases. Deep learning techniques, particularly convolutional neural networks (CNNs), have been effective in detecting and segmenting pancreatic tissues as well as differentiating between benign and malignant lesions. Deep learning algorithms have also been used to predict survival time, recurrence risk, and therapy response in pancreatic cancer patients. Radiomics approaches, extracting quantitative features from imaging modalities such as CT, MRI, and endoscopic ultrasound, have enhanced the accuracy of these deep learning models. Despite the potential of Artificial intelligence in pancreatic imaging, challenges such as legal and ethical considerations, algorithm transparency, and data security remain. This review underscores the transformative potential of Artificial intelligence in enhancing the diagnosis and treatment of pancreatic diseases, ultimately aiming to improve patient outcomes and survival rates.

The accurate detection and precise segmentation of lung nodules on computed tomography are key prerequisites for early diagnosis and appropriate treatment of lung cancer. This study was designed to compare detection and segmentation methods for pulmonary nodules using deep-learning techniques to fill methodological gaps and biases in the existing literature. This study utilized a systematic review with the Preferred Reporting Items for Systematic Reviews and Meta-Analyses guidelines, searching PubMed, Embase, Web of Science Core Collection, and the Cochrane Library databases up to May 10, 2023. The Quality Assessment of Diagnostic Accuracy Studies 2 criteria was used to assess the risk of bias and was adjusted with the Checklist for Artificial Intelligence in Medical Imaging. The study analyzed and extracted model performance, data sources, and task-focus information. After screening, we included nine studies meeting our inclusion criteria. These studies were published between 2019 and 2023 and predominantly used public datasets, with the Lung Image Database Consortium Image Collection and Image Database Resource Initiative and Lung Nodule Analysis 2016 being the most common. The studies focused on detection, segmentation, and other tasks, primarily utilizing Convolutional Neural Networks for model development. Performance evaluation covered multiple metrics, including sensitivity and the Dice coefficient. This study highlights the potential power of deep learning in lung nodule detection and segmentation. It underscores the importance of standardized data processing, code and data sharing, the value of external test datasets, and the need to balance model complexity and efficiency in future research. Deep learning demonstrates significant promise in autonomously detecting and segmenting pulmonary nodules. Future research should address methodological shortcomings and variability to enhance its clinical utility. Deep learning shows potential in the detection and segmentation of pulmonary nodules. There are methodological gaps and biases present in the existing literature. Factors such as external validation and transparency affect the clinical application.

Cancer-associated fibroblasts (CAFs) are found in primary and advanced tumours. They are primarily involved in tumour progression through complex mechanisms with other types of cells in the tumour microenvironment. However, essential fibroblasts-related genes (FRG) in bladder cancer still need to be explored, and there is a shortage of an ideal predictive model or molecular subtype for the progression and immune therapeutic assessment for bladder cancer, especially muscular-invasive bladder cancer based on the FRG. CAF-related genes of bladder cancer were identified by analysing single-cell RNA sequence datasets, and bulk transcriptome datasets and gene signatures were used to characterize them. Then, 10 types of machine learning algorithms were utilised to determine the hallmark FRG and construct the FRG index (FRGI) and subtypes. Further molecular subtypes combined with CD8+ T-cells were established to predict the prognosis and immune therapy response. Fifty-four BLCA-related FRG were screened by large-scale scRNA-sequence datasets. The machine learning algorithm established a 3-genes FRGI. High FRGI represented a worse outcome. Then, FRGI combined clinical variables to construct a nomogram, which shows high predictive performance for the prognosis of bladder cancer. Furthermore, the BLCA datasets were separated into two subtypes - fibroblast hot and cold types. In five independent BLCA cohorts, the fibroblast hot type showed worse outcomes than the cold type. Multiple cancer-related hallmark pathways are distinctively enriched in these two types. In addition, high FRGI or fibroblast hot type shows a worse immune therapeutic response. Then, four subtypes called CD8-FRG subtypes were established under the combination of FRG signature and activity of CD8+ T-cells, which turned out to be effective in predicting the prognosis and immune therapeutic response of bladder cancer in multiple independent datasets. Pathway enrichment analysis, multiple gene signatures, and epigenetic alteration characterize the CD8-FRG subtypes and provide a potential combination strategy method against bladder cancer. In summary, the authors established a novel FRGI and CD8-FRG subtype by large-scale datasets and organised analyses, which could accurately predict clinical outcomes and immune therapeutic response of BLCA after surgery.

Lysyl oxidases (LOX/LOXL1-4) are crucial for cancer progression, yet their transcriptional regulation, potential therapeutic targeting, prognostic value and involvement in immune regulation remain poorly understood. This study comprehensively evaluates LOX/LOXL expression in cancer and highlights cancer types where targeting these enzymes and developing LOX/LOXL-based prognostic models could have significant clinical relevance. We assessed the association of LOX/LOXL expression with survival and drug sensitivity via analyzing public datasets (including bulk and single-cell RNA sequencing data of six datasets from Gene Expression Omnibus (GEO), Chinese Glioma Genome Atlas (CGGA) and Cancer Genome Atlas Program (TCGA)). We performed comprehensive machine learning-based bioinformatics analyses, including unsupervised consensus clustering, a total of 10 machine-learning algorithms for prognostic prediction and the Connectivity map tool for drug sensitivity prediction. The clinical significance of the LOX/LOXL family was evaluated across 33 cancer types. Overexpression of LOX/LOXL showed a strong correlation with tumor progression and poor survival, particularly in glioma. Therefore, we developed a novel prognostic model for glioma by integrating LOX/LOXL expression and its co-expressed genes. This model was highly predictive for overall survival in glioma patients, indicating significant clinical utility in prognostic assessment. Furthermore, our analysis uncovered a distinct LOXL2-overexpressing malignant cell population in recurrent glioma, characterized by activation of collagen, laminin, and semaphorin-3 pathways, along with enhanced epithelial-mesenchymal transition. Apart from glioma, our data revealed the role of LOXL3 overexpression in macrophages and in predicting the response to immune checkpoint blockade in bladder and renal cancers. Given the pro-tumor role of LOX/LOXL genes in most analyzed cancers, we identified potential therapeutic compounds, such as the VEGFR inhibitor cediranib, to target pan-LOX/LOXL overexpression in cancer. Our study provides novel insights into the potential value of LOX/LOXL in cancer pathogenesis and treatment, and particularly its prognostic significance in glioma.

The potential of serum extracellular vesicles (EVs) as non-invasive biomarkers for diagnosing colorectal cancer (CRC) remains elusive. We employed an in-depth 4D-DIA proteomics and machine learning (ML) pipeline to identify key proteins, PF4 and AACT, for CRC diagnosis in serum EV samples from a discovery cohort of 37 cases. PF4 and AACT outperform traditional biomarkers, CEA and CA19-9, detected by ELISA in 912 individuals. Furthermore, we developed an EV-related random forest (RF) model with the highest diagnostic efficiency, achieving AUC values of 0.960 and 0.963 in the train and test sets, respectively. Notably, this model demonstrated reliable diagnostic performance for early-stage CRC and distinguishing CRC from benign colorectal diseases. Additionally, multi-omics approaches were employed to predict the functions and potential sources of serum EV-derived proteins. Collectively, our study identified the crucial proteomic signatures in serum EVs and established a promising EV-related RF model for CRC diagnosis in the clinic.

Cancer is a leading cause of death globally. Accurate cancer burden information is crucial for policy planning, but many countries do not have up-to-date cancer surveillance data. To inform global cancer-control efforts, we used the Global Burden of Diseases, Injuries, and Risk Factors Study (GBD) 2023 framework to generate and analyse estimates of cancer burden for 47 cancer types or groupings by age, sex, and 204 countries and territories from 1990 to 2023, cancer burden attributable to selected risk factors from 1990 to 2023, and forecasted cancer burden up to 2050. Cancer estimation in GBD 2023 used data from population-based cancer registration systems, vital registration systems, and verbal autopsies. Cancer mortality was estimated using ensemble models, with incidence informed by mortality estimates and mortality-to-incidence ratios (MIRs). Prevalence estimates were generated from modelled survival estimates, then multiplied by disability weights to estimate years lived with disability (YLDs). Years of life lost (YLLs) were estimated by multiplying age-specific cancer deaths by the GBD standard life expectancy at the age of death. Disability-adjusted life-years (DALYs) were calculated as the sum of YLLs and YLDs. We used the GBD 2023 comparative risk assessment framework to estimate cancer burden attributable to 44 behavioural, environmental and occupational, and metabolic risk factors. To forecast cancer burden from 2024 to 2050, we used the GBD 2023 forecasting framework, which included forecasts of relevant risk factor exposures and used Socio-demographic Index as a covariate for forecasting the proportion of each cancer not affected by these risk factors. Progress towards the UN Sustainable Development Goal (SDG) target 3.4 aim to reduce non-communicable disease mortality by a third between 2015 and 2030 was estimated for cancer. In 2023, excluding non-melanoma skin cancers, there were 18·5 million (95% uncertainty interval 16·4 to 20·7) incident cases of cancer and 10·4 million (9·65 to 10·9) deaths, contributing to 271 million (255 to 285) DALYs globally. Of these, 57·9% (56·1 to 59·8) of incident cases and 65·8% (64·3 to 67·6) of cancer deaths occurred in low-income to upper-middle-income countries based on World Bank income group classifications. Cancer was the second leading cause of deaths globally in 2023 after cardiovascular diseases. There were 4·33 million (3·85 to 4·78) risk-attributable cancer deaths globally in 2023, comprising 41·7% (37·8 to 45·4) of all cancer deaths. Risk-attributable cancer deaths increased by 72·3% (57·1 to 86·8) from 1990 to 2023, whereas overall global cancer deaths increased by 74·3% (62·2 to 86·2) over the same period. The reference forecasts (the most likely future) estimate that in 2050 there will be 30·5 million (22·9 to 38·9) cases and 18·6 million (15·6 to 21·5) deaths from cancer globally, 60·7% (41·9 to 80·6) and 74·5% (50·1 to 104·2) increases from 2024, respectively. These forecasted increases in deaths are greater in low-income and middle-income countries (90·6% [61·0 to 127·0]) compared with high-income countries (42·8% [28·3 to 58·6]). Most of these increases are likely due to demographic changes, as age-standardised death rates are forecast to change by -5·6% (-12·8 to 4·6) between 2024 and 2050 globally. Between 2015 and 2030, the probability of dying due to cancer between the ages of 30 years and 70 years was forecasted to have a relative decrease of 6·5% (3·2 to 10·3). Cancer is a major contributor to global disease burden, with increasing numbers of cases and deaths forecasted up to 2050 and a disproportionate growth in burden in countries with scarce resources. The decline in age-standardised mortality rates from cancer is encouraging but insufficient to meet the SDG target set for 2030. Effectively and sustainably addressing cancer burden globally will require comprehensive national and international efforts that consider health systems and context in the development and implementation of cancer-control strategies across the continuum of prevention, diagnosis, and treatment. Gates Foundation, St Jude Children's Research Hospital, and St Baldrick's Foundation.

Artificial intelligence (AI) is a rapidly growing field with significant implications for radiology. Acute abdominal pain is a common clinical presentation that can range from benign conditions to life-threatening emergencies. The critical nature of these situations renders emergent abdominal imaging an ideal candidate for AI applications. CT, radiographs, and ultrasound are the most common modalities for imaging evaluation of these patients. For each modality, numerous studies have assessed the performance of AI models for detecting common pathologies, such as appendicitis, bowel obstruction, and cholecystitis. The capabilities of these models range from simple classification to detailed severity assessment. This narrative review explores the evolution, trends, and challenges in AI applications for evaluating acute abdominal pathologies. We review implementations of AI for non-traumatic and traumatic abdominal pathologies, with discussion of potential clinical impact, challenges, and future directions for the technology.

Early screening is crucial in reducing the mortality of colorectal cancer (CRC). Current screening methods, including fecal occult blood tests (FOBT) and colonoscopy, are primarily limited by low patient compliance and the invasive nature of the procedures. Several advanced imaging techniques such as computed tomography (CT) and histological imaging have been integrated with artificial intelligence (AI) to enhance the detection of CRC. There are still limitations because of the challenges associated with image acquisition and the cost. Kidney, ureter, and bladder (KUB) radiograph which is inexpensive and widely used for abdominal assessments in emergency settings and shows potential for detecting CRC when enhanced using advanced techniques. This study aimed to develop a deep learning model (DLM) to detect CRC using KUB radiographs. This retrospective study was conducted using data from the Tri-Service General Hospital (TSGH) between January 2011 and December 2020, including patients with at least one KUB radiograph. Patients were divided into development (n = 28,055), tuning (n = 11,234), and internal validation (n = 16,875) sets. An additional 15,876 patients were collected from a community hospital as the external validation set. A 121-layer DenseNet convolutional network was trained to classify KUB images for CRC detection. The model performance was evaluated using receiver operating characteristic curves, with sensitivity, specificity, and area under the curve (AUC) as metrics. The AUC, sensitivity, and specificity of the DLM in the internal and external validation sets achieved 0.738, 61.3%, and 74.4%, as well as 0.656, 47.7%, and 72.9%, respectively. The model performed better for high-grade CRC, with AUCs of 0.744 and 0.674 in the internal and external sets, respectively. Stratified analysis showed superior performance in females aged 55-64 with high-grade cancers. AI-positive predictions were associated with a higher long-term risk of all-cause mortality in both validation cohorts. AI-enhanced KUB X-ray analysis can enhance CRC screening coverage and effectiveness, providing a cost-effective alternative to traditional methods. Further prospective studies are necessary to validate these findings and fully integrate this technology into clinical practice.

Multiplexed immunofluorescence (mIF) staining, such as CODEX and MIBI, holds significant clinical value for various fields, such as disease diagnosis, biological research, and drug development. However, these techniques are often hindered by high time and cost requirements. Here we present a Multimodal-Attention-based virtual mIF Staining (MAS) system that utilises a deep learning model to extract potential antibody-related features from dual-modal non-antibody-stained fluorescence imaging, specifically autofluorescence (AF) and DAPI imaging. The MAS system simultaneously generates predictions of mIF with multiple survival-associated biomarkers in gastric cancer using self- and multi-attention learning mechanisms. Experimental results with 180 pathological slides from 94 patients with gastric cancer demonstrate the efficiency and consistent performance of the MAS system in both cancer and noncancer gastric tissues. Furthermore, we showcase the prognostic accuracy of the virtual mIF images of seven gastric cancer related biomarkers, including CD3, CD20, FOXP3, PD1, CD8, CD163, and PD-L1, which is comparable to those obtained from the standard mIF staining. The MAS system rapidly generates reliable multiplexed staining, greatly reducing the cost of mIF and improving clinical workflow. Stanford 2022 HAI Seed Grant; National Institutes of Health 1R01CA256890.

Exocrine and endocrine pancreas are interconnected anatomically and functionally, with vasculature facilitating bidirectional communication. Our understanding of this network remains limited, largely due to two-dimensional histology and missing combination with three-dimensional imaging. In this study, a multiscale 3D-imaging process was used to analyze a porcine pancreas. Clinical computed tomography, digital volume tomography, micro-computed tomography and Synchrotron-based propagation-based imaging were applied consecutively. Fields of view correlated inversely with attainable resolution from a whole organism level down to capillary structures with a voxel edge length of 2.0 µm. Segmented vascular networks from 3D-imaging data were correlated with tissue sections stained by immunohistochemistry and revealed highly vascularized regions to be intra-islet capillaries of islets of Langerhans. Generated 3D-datasets allowed for three-dimensional qualitative and quantitative organ and vessel structure analysis. Beyond this study, the method shows potential for application across a wide range of patho-morphology analyses and might possibly provide microstructural blueprints for biotissue engineering.

Postoperative recurrence risk for pediatric low-grade gliomas (pLGGs) is challenging to predict by conventional clinical, radiographic, and genomic factors. We investigated if deep learning (DL) of magnetic resonance imaging (MRI) tumor features could improve postoperative pLGG risk stratification. We used a pretrained DL tool designed for pLGG segmentation to extract pLGG imaging features from preoperative T2-weighted MRI from patients who underwent surgery (DL-MRI features). Patients were pooled from 2 institutions: Dana Farber/Boston Children's Hospital (DF/BCH) and the Children's Brain Tumor Network (CBTN). We trained 3 DL logistic hazard models to predict postoperative event-free survival (EFS) probabilities with (1) clinical features, (2) DL-MRI features, and (3) multimodal (clinical and DL-MRI features). We evaluated the models with a time-dependent Concordance Index (Ctd) and risk group stratification with Kaplan-Meier plots and log-rank tests. We developed an automated pipeline integrating pLGG segmentation and EFS prediction with the best model. Of the 396 patients analyzed (median follow-up: 85 months, range: 1.5-329 months), 214 (54%) underwent gross total resection and 110 (28%) recurred. The multimodal model improved EFS prediction compared to the DL-MRI and clinical models (Ctd: 0.85 (95% CI: 0.81-0.93), 0.79 (95% CI: 0.70-0.88), and 0.72 (95% CI: 0.57-0.77), respectively). The multimodal model improved risk-group stratification (3-year EFS for predicted high-risk: 31% versus low-risk: 92%, P < .0001). DL extracts imaging features that can inform postoperative recurrence prediction for pLGG. Multimodal DL improves postoperative risk stratification for pLGG and may guide postoperative decision-making. Larger, multicenter training data may be needed to improve model generalizability.

Parkinson's disease (PD) is an incurable neurological disorder that often begins insidiously with sleep disturbances and somatic symptoms, progressing to whole-body motor and cognitive symptoms1-5. Dysfunction of the somato-cognitive action network (SCAN)-which is thought to control action execution6,7 by coordinating arousal, organ physiology and whole-body motor plans with behavioural motivation-is a potential contributor to the diverse clinical manifestations of PD. To investigate the role of the SCAN in PD pathophysiology and treatments (medications, deep-brain stimulation (DBS), transcranial magnetic stimulation (TMS) and MRI-guided focused ultrasound stimulation (MRgFUS)), we built a large (n = 863), multimodal, multi-intervention clinical imaging dataset. Resting-state functional connectivity revealed that the substantia nigra and all PD DBS targets (subthalamic nucleus, globus pallidus and ventral intermediate thalamus) are selectively connected to the SCAN rather than to effector-specific motor regions. Importantly, PD was characterized by specific hyperconnectivity between the SCAN and the subcortex. We therefore followed six PD cohorts undergoing DBS, TMS, MRgFUS and levodopa therapy using precision resting-state functional connectivity and electrocorticography recording. Efficacious treatments reduced SCAN-to-subcortex hyperconnectivity. Targeting the SCAN instead of effector regions doubled the efficacy of TMS treatments. Focused ultrasound treatment benefits increased when the target was closer to the thalamic SCAN sweet spot. Thus, SCAN hyperconnectivity is central to PD pathophysiology and its alleviation is a hallmark of successful neuromodulation. Targeting functionally defined subcortical SCAN nodes may improve existing therapies (DBS, MRgFUS), whereas cortical SCAN targets offer effective non-invasive or minimally invasive neuromodulation for PD.

Magnetic resonance imaging (MRI) is crucial for diagnosing and monitoring of multiple sclerosis (MS) as it is used to assess lesions in the brain and spinal cord. However, in real-world clinical settings, MRI scans are often acquired with thick slices, limiting their utility for automated quantitative analyses. This work presents a single-image super-resolution (SR) reconstruction framework that leverages SR convolutional neural networks (CNN) to enhance the through-plane resolution of structural MRI in people with MS (PwMS). Our strategy involves the supervised fine-tuning of CNN architectures, guided by a content loss function that promotes perceptual quality, as well as reconstruction accuracy, to recover high-level image features. Extensive evaluation with MRI data of PwMS shows that our SR strategy leads to more accurate MRI reconstructions than competing methods. Furthermore, it improves lesion segmentation on low-resolution MRI, approaching the performance achievable with high-resolution images. Results demonstrate the potential of our SR framework to facilitate the use of low-resolution retrospective MRI from real-world clinical settings to investigate quantitative image-based biomarkers of MS.

This study evaluated the performance of a multi-stage Segmentation Residual Network (SegResNet)-based deep learning (DL) model for the automatic segmentation of cone-beam computed tomography (CBCT) images of patients with stage III and IV periodontitis. Seventy pre-processed CBCT scans from patients undergoing periodontal rehabilitation were used for training and validation. The model was tested on 10 CBCT scans independent from the training dataset by comparing results with semi-automatic (SA) segmentations. Segmentation accuracy was assessed using the Dice similarity coefficient (DSC), Intersection over Union (IoU), and Hausdorff distance 95th percentile (HD95). Linear periodontal measurements were performed on four tooth surfaces to assess the validity of the DL segmentation in the periodontal region. The DL model achieved a mean DSC of 0.9650 ± 0.0097, with an IoU of 0.9340 ± 0.0180 and HD95 of 0.4820 mm ± 0.1269 mm, showing strong agreement with SA segmentation. Linear measurements revealed high statistical correlations between the mesial, distal, and lingual surfaces, with intraclass correlation coefficients (ICC) of 0.9442 (p < 0.0001), 0.9232 (p < 0.0001), and 0.9598(p < 0.0001), respectively, while buccal measurements revealed lower consistency, with an ICC of 0.7481 (p < 0.0001). The DL method reduced the segmentation time by 47 times compared to the SA method. Acquired 3D models may enable precise treatment planning in cases where conventional diagnostic modalities are insufficient. However, the robustness of the model must be increased to improve its general reliability and consistency at the buccal aspect of the periodontal region. This study presents a DL model for the CBCT-based segmentation of periodontal defects, demonstrating high accuracy and a 47-fold time reduction compared to SA methods, thus improving the feasibility of 3D diagnostics for advanced periodontitis.

The information-processing capability of the brain's cellular network depends on the physical wiring pattern between neurons and their molecular and functional characteristics. Mapping neurons and resolving their individual synaptic connections can be achieved by volumetric imaging at nanoscale resolution1,2 with dense cellular labelling. Light microscopy is uniquely positioned to visualize specific molecules, but dense, synapse-level circuit reconstruction by light microscopy has been out of reach, owing to limitations in resolution, contrast and volumetric imaging capability. Here we describe light-microscopy-based connectomics (LICONN). We integrated specifically engineered hydrogel embedding and expansion with comprehensive deep-learning-based segmentation and analysis of connectivity, thereby directly incorporating molecular information into synapse-level reconstructions of brain tissue. LICONN will allow synapse-level phenotyping of brain tissue in biological experiments in a readily adoptable manner.