搜索结果：Journal of imaging informatics in medicine

The objective of this study is to demonstrate an end-to-end approach for operationalizing the Findable, Accessible, Interoperable, and Reusable (FAIR) principles in multimodal medical imaging research using standardized data models and reproducible computational workflows, illustrated by reproducing the design and directional findings of a published Alzheimer's disease (AD) imaging study. Clinical and imaging data from the Alzheimer's Disease Neuroimaging Initiative (ADNI-4) were harmonized within the Observational Medical Outcomes Partnership Common Data Model and its Medical Imaging extension (OMOP MI-CDM). MRI acquisition metadata (DICOM) were extracted and mapped to standardized concepts, while clinical variables were integrated via reproducible extract-transform-load processes. Interoperable phenotypes were defined using OHDSI tools. Hippocampal volumes were derived from T1-weighted MRI using a fully automated machine learning segmentation pipeline (OpenMap-T1). Imaging attributes and derived measurements were stored as structured, provenance-preserving records in OMOP MI-CDM. We replicated a reference study evaluating hippocampal volume differences across AD, mild cognitive impairment (MCI), and cognitively normal controls, stratified by age and sex. We included 289 participants and 545 MRI studies. Across age- and sex-stratified cohorts, mean hippocampal volumes showed consistent directional reductions in AD compared with controls, with intermediate values in MCI, matching trends reported in the replication study. FAIR principles can be operationalized across the full imaging research pipeline using OMOP MI-CDM and automated analysis workflows. This framework enables transparent cohort definition, reproducible image processing, and interoperable reuse of machine learning-derived imaging features to support scalable validation and reproducible multimodal imaging research.

Efficiently identifying and retrieving imaging studies to create research cohorts is often hindered by the segregated storage of medical imaging data and radiological reports. To address this, we developed PACScrawler, an open-source tool designed to streamline cohort assembly by enabling efficient search and access to imaging studies across Picture Archiving and Communication Systems (PACS) and Radiology Information Systems (RIS). The system features a backend crawler module that regularly collects DICOM and RIS metadata and radiology reports from newly acquired studies and stores them in a local Apache Solr/Lucene index. A web-based interface enables multi-parameter search across the indexed database, and a retrieval module downloads the corresponding imaging data on demand. Between 2012 and 2024, PACScrawler indexed 2,327,317 examinations from 316,310 patients at our institution. In our local setup, routine indexing of 1079 examinations on a representative day required 26.5 min, corresponding to a median of 1.4 s per study. In a batch retrieval benchmark, ten head MRI examinations comprising 325 series were retrieved in 16 min 08 s using PACScrawler compared with 31 min 45 s by manual sequential PACS export. For LLM-assisted query construction, 17 of 20 generated Solr queries were usable with no or limited manual adjustment. As an illustrative use case, the index enabled longitudinal monitoring of local imaging volumes and examination complexity. PACScrawler lowers the technical barrier for clinical imaging research by providing an open-source, regularly updated local search index for selected PACS/RIS metadata and radiology reports. It supports retrospective cohort identification, on-demand retrieval of imaging data, and longitudinal monitoring of institutional imaging trends while routine indexing is limited to metadata and report text. As a DICOM-based tool, PACScrawler may be adapted to other clinical environments after local PACS/RIS integration and validation; source code is available at https://github.com/pacs-ris-crawler/pacs-ris-crawler .

This study aims to develop and validate an MRI-based image-language collaborative framework for differentiating tuberculous spondylitis (TB) from pyogenic spondylitis (PS) at the patient level. This retrospective two-center study included 347 patients with microbiologically or histopathologically confirmed spinal infection (218 TB, 129 PS) who underwent sagittal T2 weighted MRI between June 2021 and June 2025. An enhanced YOLO model was used to localize infection-related regions of interest (ROIs), which were subsequently reviewed by radiologists and used as semi-automated standardized inputs. Patient-level diagnostic performance was evaluated across ROI-guided supervised classifiers (ConvNeXt and Swin Transformer), whole-image large language model (LLM) zero-shot and few-shot inference, and ROI-guided LLM few-shot inference. Multivariable logistic regression was performed to identify independent imaging features. Vertebral collapse (OR, 3.713; p&#xa0;< 0.001) and marked disc involvement (OR, 0.209; p < 0.001) were independent imaging features for differentiating TB from PS. The lesion detector achieved an mAP@0.5 of 0.97. After patient-level probability pooling and validation-set thresholding, ConvNeXt achieved an AUC of 0.762 and an accuracy of 62.9%, whereas Swin Transformer achieved an AUC of 0.804 and an accuracy of 71.4% on the 35-patient independent test set. Whole-image zero-shot LLM inference showed limited diagnostic value, improving with few-shot prompting. On the same 35-patient test set, ROI guided 10-shot LLM achieved 82.9% accuracy, 81.8% sensitivity, 84.6% specificity, 90.0% precision, and 85.7% F1 score, showing numerically more balanced patient-level performance than whole-image inference. Because of the limited test-set size, these performance estimates should be interpreted as exploratory. ROI-guided constrained LLM reasoning provides stable and interpretable patient-level differentiation of spinal infection on MRI. By linking diagnostic output to traceable ROI-level evidence, this framework may serve as a radiologist-supervised second-reader tool for difficult TB/PS cases.

Automated analysis of lumbar spine MRI remains constrained by the absence of a comprehensive dataset integrating pixel-level anatomy, quantitative markings, and radiological reports within a unified framework. We present LSMA-PQR, a clinically validated lumbar spine MRI dataset comprising N = 515 patients with dual-plane, dual-sequence imaging (axial and sagittal T1/T2-weighted; 4120 images) and comprehensive multi-modal annotations. Pixel-level segmentation masks for seven anatomical structures are provided in multiple interoperable formats with verified cross-format spatial consistency (mean IoU = 0.994 &#xb1; 0.008 ). Beyond anatomy, the dataset integrates quantitative disc markings including 1545 Pfirrmann degeneration grades (L3-L4, L4-L5, L5-S1) and intervertebral disc height measurements with pixel-coordinate provenance, as well as structured radiological reports derived from 515 free-text documents through systematic natural language processing that identified and corrected 340 transcription errors across 59% of reports. Clinical validation of the dataset reveals peak degeneration at L5-S1 (71%), a 40% prevalence of severe pathology (Pfirrmann Grade 4-5), and a moderate-to-strong correlation between clinical severity and expert-assigned grades. Baseline experiments on three downstream tasks, multi-plane segmentation (mean Dice = 0.946), disc-aware Pfirrmann grading (quadratic weighted &#x3ba; = 0.51), and structured clinical findings extraction (weighted F1 = 0.59), demonstrate the dataset's utility for training and evaluating automated lumbar spine analysis systems. LSMA-PQR, available at Mendeley Data , enables previously infeasible research including report-guided supervision, multi-metric cross-validation and integrated modeling of lumbar spine anatomy, degeneration and clinical findings. Data and documentation are released under a CC BY 4.0 license to accelerate translational spine imaging research.

Intracranial hemorrhage (ICH) is a time-critical neurologic emergency where delayed diagnosis can worsen outcomes. Non-contrast head CT is the first-line imaging modality, and explainable AI tools may enhance rapid detection and subtype classification. This study presents an explainable deep learning framework for multi-label ICH detection using DICOM-native multi-window CT data with leakage-aware validation. A retrospective secondary analysis was performed using the publicly available training dataset from the RSNA 2019 Intracranial Hemorrhage Detection Challenge, comprising 752,803 axial CT images from 18,938 patients and 21,744 studies. Images were processed directly from DICOM, converted to Hounsfield units, and encoded into three channels (brain, subdural, bone windows). Three independent ResNet34-based multi-label models were trained using weighted binary cross-entropy within a threefold patient- and study-level cross-validation framework. The fold-specific models were evaluated separately, without model merging or ensembling, and performance was summarized across folds. Performance was assessed using ROC AUC and PR AUC, along with threshold-dependent metrics (precision, sensitivity, F1 score, and specificity) and the Brier score after class-specific threshold optimization. Explainability was evaluated using Grad-CAM, Integrated Gradients, occlusion sensitivity, quantitative cross-method agreement, window ablation, and deletion-insertion faithfulness analyses. The model showed strong and consistent discrimination. For any hemorrhage, ROC AUC ranged from 0.973 to 0.974 and PR AUC from 0.894 to 0.899. Macro ROC AUC ranged from 0.968 to 0.971 and micro ROC AUC from 0.976 to 0.978. Mean macro precision, sensitivity, and F1 score were 0.640, 0.670, and 0.646, respectively, with a specificity of 0.984 and a Brier score of 0.054. Best F1 scores were observed for any hemorrhage (0.822), intraventricular (0.788), and intraparenchymal hemorrhage (0.767), while epidural hemorrhage remained low (0.203). Across 72 class-specific explanations from 56 CT images, spatial agreement among attribution methods was moderate, window ablation indicated substantial joint dependence on all three CT windows, and the deletion and insertion areas under the curve were 0.185&#x2009;&#xb1;&#x2009;0.147 and 0.882&#x2009;&#xb1;&#x2009;0.110, respectively, supporting a functional association between Grad-CAM-highlighted regions and model confidence. This DICOM-native, explainable approach achieved robust ICH detection, though performance for rare subtypes remains limited due to class imbalance.

The purpose of this study is to develop and retrospectively validate a deep learning system for the classification and anatomically interpretable localization of six acute abdominal emergencies on CT using multi-window Hounsfield Unit (HU) encoding. A publicly available national teleradiology dataset of 1274 patients (42,922 bounding box annotations) was used for training and internal validation (896/189/189 patient-level split). Each CT slice was encoded into three diagnostic HU windows (soft tissue, bone/stone, angio/liver). A YOLOv11-Large model with a stride-4 (P2) head was trained at 1280 &#xd7; 1280 . Localization was evaluated using the clinical nine-region abdominal grid, and specificity was assessed in 80 target-negative patients. External validation used a radiologist-adjudicated 280-patient Stanford Merlin cohort (US), with model weights and thresholds applied without modification. Internal macro AUROC was 0.941, and macro F1 was 76.1%. Nine-region localization accuracy was 99.5% among detected cases and 90.9% including missed detections. Specificity in the target-negative cohort was 86.2%. On the external Stanford Merlin cohort, macro AUROC was 0.879 with all six classes &#x2265; 0.80 ; AAA reached F1&#xa0;0.889 at frozen thresholds; macro F1 was 0.545, increasing to 0.648 after recalibration. A multi-window CT detection model classified six acute abdominal emergencies with high discrimination on internal testing and preserved moderate-to-high discrimination on an external Stanford Merlin cohort. Region-level evaluation using the nine-region abdominal grid provided a clinically interpretable localization endpoint complementary to conventional detection metrics. However, variable class-level F1, reduced external operating-point performance, and retrospective design indicate that multisite prospective validation and site-specific threshold calibration are required before clinical deployment.

Large language models (LLMs) are increasingly explored in radiology, yet concerns persist regarding hallucination and lack of factual grounding. Retrieval-augmented generation (RAG) seeks to address these limitations by coupling generative models with external knowledge retrieval. We conducted a scoping review to characterize how RAG systems have been applied in radiology and medical imaging. A systematic search of PubMed, Embase, Scopus, IEEE Xplore, and arXiv identified 45 studies implementing RAG-based approaches in radiology-related tasks. In terms of clinical tasks, RAG was most commonly applied to radiology report generation and question answering. Dense retrieval strategies predominated, while sparse, hybrid and proprietary retrieval approaches were less frequent. External knowledge sources most frequently comprised biomedical literature databases and clinical guidelines. Applications were heavily skewed toward chest radiography and X-ray-based tasks, with relatively few studies addressing CT, MRI, PET, ultrasound, or under-represented subspecialties such as pediatric radiology and neuroradiology. Most comparative studies reported task-specific performance gains with RAG over non-retrieval baselines, and a small subset reported performance comparable to trained radiologists or state-of-the-art models. However, hallucinations and errors persisted, and heterogeneity across studies limited the generalizability of these findings. Evaluation practices largely relied on automated accuracy or text-overlap metrics, with limited use of standardized expert evaluation and minimal assessment of safety, bias, computational efficiency, or clinical utility. Overall, while RAG shows promise for improving factual grounding in radiology AI, current evaluation paradigms likely overestimate real-world clinical readiness. Future work should prioritize retrieval quality, clinically grounded evaluation, safety-critical error analysis, bias assessment, and deployment-relevant efficiency metrics to enable responsible clinical translation.

Deep learning (DL) is increasingly applied to automate brain tumor classification from magnetic resonance imaging (MRI), yet meaningful clinical deployment remains limited by tumor heterogeneity, dataset bias, and incomplete tumor-specific validation. This systematic review synthesizes developments from 2016 to 2025 in advanced DL-based MRI brain tumor classification, with a specific emphasis on meningioma-focused classification and subtyping, given their persistent underrepresentation in AI research. Fifty-six eligible studies were analyzed and organized into five methodological categories: transformer-based models; transformer-based feature extraction pipelines; attention-enhanced convolutional neural networks (CNNs); federated learning approaches; and emerging or unconventional DL strategies. Among studies reporting class-wise metrics (n&#x2009;=&#x2009;19), attention-enhanced CNNs and hybrid CNN-transformer architectures showed high overall accuracy with fewer extreme drops in meningioma performance (F1-score, 89.0% to 99.0%) than end-to-end transformers (F1-score, 79.0% to 97.0%), despite the latter achieving high peak accuracies. Our analysis also showed a marked gap between general tumor categorization and clinically actionable subtyping or grade classification, particularly for meningiomas, reflecting both limited meningioma-targeted tasks and reduced transparency in per-class reporting. Study design and reporting practices limited cross-study comparability, with heavy reliance on a small number of publicly available datasets, frequent class imbalance disadvantaging meningioma representation, and approximately 60% of studies not specifying MRI sequence details. Although predictive performance improved and some studies incorporated interpretability or&#xa0;clinical decision-support&#xa0;components, reporting remained sparse, reinforcing the translational gap between methodological progress and deployment readiness. Finally, publication activity accelerated sharply after 2023 but remained geographically concentrated, raising concerns about representativeness. Our findings call for standardized per-class reporting, greater&#xa0;diverse datasets, interpretability components,&#xa0;and clinically aligned meningioma evaluations to support effective translation into practice.

Microvascular invasion (MVI) represents a critical prognostic determinant in hepatocellular carcinoma (HCC), yet preoperative prediction remains challenging. MRI-based radiomics offers promise for noninvasive MVI stratification through quantitative feature extraction from intratumoral and peritumoral regions, with growing interest in integrating both compartments to capture tumor-host interactions. A comprehensive systematic literature search of PubMed, Embase, Scopus, and Web of Science through October 29, 2025, identified studies developing MRI-based radiomics or deep learning models for MVI prediction in HCC with histopathological reference standards. Random-effects meta-analysis synthesized diagnostic accuracy metrics. Risk of bias and methodological quality were assessed using QUADAS-2 and RQS-2. Thirteen retrospective studies (N&#x2009;=&#x2009;3173 patients) met inclusion criteria. Combined intratumoral-peritumoral models demonstrated the highest diagnostic accuracy with pooled area under the receiver operating characteristic curve (AUC) of 0.83 (95% CI: 0.80-0.87). Peritumoral-only models (7 studies, n&#x2009;=&#x2009;473) yielded AUC&#x2009;=&#x2009;0.78, comparable to intratumoral-only approaches (8 studies, n&#x2009;=&#x2009;633; AUC&#x2009;=&#x2009;0.78). The incremental diagnostic benefit of combined models over peritumoral-only (delta AUC&#x2009;=&#x2009;0.05, p&#x2009;=&#x2009;0.110) and intratumoral-only models (delta AUC&#x2009;=&#x2009;0.05, p&#x2009;=&#x2009;0.097) was modest and not statistically significant. Peritumoral ring width did not significantly modify performance (p&#x2009;=&#x2009;0.828). Subgroup analysis revealed that methodologically rigorous studies employing external validation achieved lower pooled AUC compared to internally validated investigations, suggesting optimism bias in reported metrics. MRI radiomics enables promising noninvasive MVI prediction in HCC, though combined intratumoral-peritumoral approaches provide limited incremental benefit over single-compartment models. Prospective multicenter external validation with standardized preprocessing and nested cross-validation protocols remain essential for clinical translation and evidence credibility.

Multimodal large language models (LLMs) offer significant potential for image-based diagnostics, yet their reliability in routine clinical microbiology workflows, specifically Gram stain interpretation, remains insufficiently characterized. In this prospective diagnostic accuracy study, 100 Gram-stained blood culture smear images from routine laboratory workflows were evaluated by three multimodal LLMs (ChatGPT-4o, Gemini 2.5 Flash, and Claude Opus). Models were assessed using standardized zero-shot prompts across two independent runs to evaluate intramodel consistency. Performance was benchmarked against three expert medical microbiologists and automated organism identification (reference standard) across Gram type, major morphology, combined classification, and fine-grained sub-details. Expert microbiologists achieved 100% concordance with the reference standard for the predefined primary categories of Gram type and major morphology. LLM accuracy was high for Gram-type classification (95-98%) but lower for cellular morphology (84-85%), resulting in combined accuracies of 82-84%. No significant inter-model differences were observed in primary tasks. Intramodel consistency was high across runs (Cohen's &#x3ba;&#x2009;=&#x2009;0.73-0.96). However, performance declined for morphological sub-details, where ChatGPT and Gemini significantly outperformed Claude. Organism-level accuracy was high for Gram-positive cocci and Gram-negative bacilli but substantially lower for Gram-positive bacilli and yeast. Multimodal large language models show promising baseline performance in Gram-stained blood culture interpretation, particularly for Gram type and major morphology. However, reduced accuracy in combined interpretation, fine-grained morphology, and less common organism classes highlights current limitations of out-of-the-box models. These findings support the use of multimodal LLMs as expert-supervised decision-support tools rather than standalone systems and emphasize the need for task-specific optimization and multicenter validation before routine clinical implementation.

Age-related cataract is a leading cause of visual impairment among older adults, significantly affecting quality of life and independence. Accurate severity grading is essential for determining optimal surgical timing and improving clinical outcomes; however, conventional slit-lamp examination relies heavily on subjective assessment, resulting in considerable inter-observer variability. In this study, we propose a progressive ordinal attention network (POA-Net) for automated four-level cataract classification. The proposed method is built on a convolutional neural network backbone and introduces a progressive ordinal cumulative attention mechanism to explicitly capture the continuous progression of disease severity. A spatial gating module is further incorporated to suppress background interference and enhance feature representation of the lens region. In addition, an ordinal prototype projection head with a topology-constrained learning strategy is designed to enforce structured feature distributions consistent with clinical knowledge. Experimental results demonstrate that POA-Net achieves an accuracy of 89.20% and a weighted kappa coefficient of 93.82%, outperforming several state-of-the-art convolutional and Transformer-based methods. The model shows improved capability in distinguishing adjacent severity levels and exhibits strong agreement with clinical annotations. These findings indicate that the proposed approach provides an accurate, consistent, and interpretable solution for automated cataract grading, with potential applications in large-scale screening and long-term monitoring, particularly in resource-limited settings.

Medical imaging plays a significant role in diagnosis and treatment planning, with significant efforts focused on training machine learning (ML) algorithms to perform detection and classification automatically. However, concerns regarding patient privacy have prompted the need for robust deidentification of the data. In this paper, we consider a general scenario in which the data are entirely anonymized by removing the metadata due to privacy concerns. In several circumstances, efficient training of an arbitrary downstream ML model requires some prior information that is only accessible through metadata. To address this need, we propose a novel approach for automated metadata prediction from fully anonymized CT and MRI images to enable efficient training of the downstream model. Our method accurately identifies the imaging modality (CT or MRI) and anatomical region (heart, brain, or liver) directly from image data. Moreover, in case of MRI data, it allows for contrast classification into T1 and T2. Our framework employs a set of machine learning as well as deterministic techniques to perform these tasks, achieving high accuracy rates in distinguishing between CT and MRI scans and localizing anatomical regions, as well as classification of MRI imaging protocol, T1 and T2. By estimating technical metadata from anonymized data, our approach successfully combines multiple tasks into a unified framework designed for privacy-preserving workflows. We demonstrate the effectiveness and practicality of our method through systematic experiments on medical imaging datasets. On held-out test sets, the framework achieved 100% accuracy for CT/MRI modality detection; 99.2% accuracy for anatomical region classification across brain, heart, and liver volumes (125/126 correctly classified); and 99.8% accuracy for MRI T1/T2 protocol classification (500/501 correctly classified). The proposed framework represents a new direction toward full privacy protection in medical imaging with no impact on the selection of appropriate image processing frameworks in the downstream tasks.

The use of artificial intelligence (AI) models to assist clinical decisions for diagnosis and prediction has been shown to improve patient outcomes and clinical decision-making. However, the performance of an AI model deployed in a clinical setting can vary over time or between initial evaluations and clinical use because of data drift. Monitoring the performance of a deployed AI model may help to (i) ensure the AI model performs as expected in the clinical environment and (ii) detect performance deviations and alert stakeholders to these deviations. In this study, we investigate how a change in the performance of an AI model caused by an abrupt data drift can be detected using a cumulative sum (CUSUM) control chart. We demonstrate the use of CUSUM for computer-aided breast cancer detection using data from the publicly available Emory Breast Imaging Dataset (EMBED). Our results indicate that, when the magnitude of the drift is 1.5 times the standard deviation of the performance metric, CUSUM is able to detect changes in the test negativity rate of an AI model within an average of 5 days following the onset of performance drift, with a long duration ( &#x223c; 293 days) between false alarms. Our results also indicate that when the average number of samples is 40/day, leading to a reasonable standard deviation for the performance metric, almost no false alarms are expected in a period of 60 days, with the choice of CUSUM parameters that lead to a small true alarm detection delay. We also demonstrate that CUSUM can be applied for monitoring the performance of AI models even when ground-truth labels are not available. We evaluate the robustness of the CUSUM chart to non-normal distributions and show that when tuned to detect relatively small parameter shifts, the CUSUM chart can be robust to non-normal distributions. We demonstrate that the sensitivity of CUSUM can be controlled to balance the mean time between false alarms and the delay in detecting a true change in performance. We conclude that CUSUM, which is a well-studied statistical process control method, can easily be adapted for monitoring the performance of AI models targeted for assisting in medical diagnosis.

Accurate assessment of cervical spine alignment is essential for diagnosing and monitoring cervical spine disorders. However, manual interpretation of cervical spine X-ray images is time-consuming and prone to variability. To address this, the study proposes an automated web-based prototype that uses a convolutional neural network (CNN) and a rule-based Cobb angle classification for cervical spinal conditions (lordosis, kyphosis, and straight neck). A stratified fivefold cross-validation (CV) framework with separate train, validation, and test splits per fold, incorporating transfer learning (TL), fine-tuning (FT), and hyperparameter optimisation (HPO) using Optuna, was employed. Two pre-trained CNNs, ResNet152V2 and ConvNeXt Tiny, were evaluated using the same data splits. Results showed that ResNet152V2 outperformed ConvNeXt Tiny, achieving a mean accuracy of 86.0&#x2009;&#xb1;&#x2009;1.21% (95% bootstrap confidence intervals [CI] 84.2%-87.9%), demonstrating a greater stability with a lower standard deviation (SD) and a narrower CI. Both the image-level paired permutation test and the McNemar test showed a statistically significant performance difference of 4.47%. Despite ResNet152V2's heavier architecture, higher GPU latency, and lower GPU throughput, the practical processing time difference is minor because the computation runs on a backend server rather than on a resource-constrained mobile platform. Therefore, the optimised ResNet152V2 model was selected for deployment via a Gradio web interface hosted on Hugging Face. In addition, the CNN-based classification significantly outperformed the Cobb angle-based classification (96.67% vs 88.33% accuracy). A 95% bootstrap CI with 10,000 resamples showed a significant improvement in CNN-based classification over the rule-based classification. A right-tailed test yielded a p-value of 0.0292 at the 0.05 significance level, consistent with 97.08% of bootstrap samples showing a positive difference in accuracy. Ultimately, this study demonstrates the feasibility of integrating an optimised deep learning model and Cobb-angle classification into a real-time cervical spine research prototype. Future work will focus on larger datasets and lightweight model designs to enhance its applicability.

The objective was to develop and validate an explainable artificial intelligence (AI)-based multimodal approach for preoperative risk stratification of central lymph node metastasis (CLNM) in papillary thyroid carcinoma (PTC) and to evaluate its role in supporting radiologist decision-making. This multicenter retrospective study enrolled patients with pathologically confirmed PTC from four hospitals. Preoperative two-dimensional ultrasound, strain elastography, shear-wave elastography, and clinical variables were integrated to develop a multimodal predictive model. Model interpretability was achieved using SHapley Additive exPlanations (SHAP) to provide feature-level explanations supporting clinical interpretation. To assess clinical usability, a controlled reader study was conducted in which six radiologists with varying experience independently evaluated cases under three conditions: without AI assistance, with basic AI assistance (probability output only), and with explainable AI assistance (visualized feature-level contributions). Model performance was evaluated using the area under the receiver operating characteristic curve (AUC), and reader performance was assessed using paired statistical comparisons and interreader agreement analysis. A total of 428 patients (mean age, 44&#xa0;years&#x2009;&#xb1;&#x2009;12; 369 women) with 508 PTC nodules were included, of whom 225 (44.3%) had CLNM. The multimodal model achieved AUCs of 0.975, 0.917, and 0.844 in the training, validation, and external test cohorts, respectively, outperforming single-modality and simplified fusion approaches (p&#x2009;<&#x2009;0.05). SHAP identified age, texture-derived radiomic features, and elastography-derived stiffness-related features as key contributors. In the reader study, explainable AI assistance significantly improved diagnostic accuracy across all experience levels, increased diagnostic confidence, and raised human-AI agreement to substantial or almost-perfect levels. An explainable AI-based multimodal approach enables accurate preoperative risk stratification of CLNM in PTC and improves radiologist diagnostic performance, with potential to support clinical decision-making within radiology workflows.

Optical Coherence Tomography Angiography (OCTA) enables high-resolution visualization of the retinal vascular network. However, retinal vessels exhibit complex multi-scale and multi-directional patterns, particularly in thin vessels, bifurcations, and crossings. Accurate characterization of these structures is important for maintaining vascular connectivity and structural integrity, which are critical for reliable retinal vascular analysis and disease assessment. Existing methods often struggle to simultaneously capture directional continuity, long-range contextual dependencies, and fine-grained structural details, resulting in vessel discontinuities or mis-segmentation in small vessels and complex junction regions. To address these challenges, we propose an Adaptive Directional Feature Fusion Network (ADF-Net) for retinal vessel segmentation in OCTA images. Specifically, ADF-Net adopts a dual-branch encoder that integrates convolution-based local feature extraction and Swin Transformer-based global context modeling. Within this framework, a directional-aware multi-scale module is introduced with adaptive aggregation over fixed directional responses, modeling vascular structures in a 12-direction orientation space spanning 360 degrees. A direction-wise attention mechanism adaptively learns the relative importance of responses from different orientations and aggregates them accordingly, improving directional continuity in bifurcation and crossing regions. Multi-scale convolutions are incorporated to capture vessels of varying calibers, and an Edge Complement Module is introduced to explicitly refine vessel boundaries and reduce structural discontinuities. Experimental results on the OCTA-500 and ROSE-1 datasets demonstrate that the proposed method achieves competitive performance with several state-of-the-art approaches in retinal vessel segmentation. Specifically, on the OCTA-3&#xa0;M, OCTA-6&#xa0;M, and ROSE-1 datasets, ADF-Net achieves Dice coefficients of 91.51%, 88.88%, and 88.03%, and Intersection over Union (IoU) scores of 84.42%, 80.07%, and 78.61%, respectively. The results demonstrate that the proposed method effectively preserves vascular continuity and structural integrity for retinal vascular analysis and disease assessment. The source code of the proposed ADF-Net is publicly available at https://github.com/suxinli41-lgtm/ADF-Net .

Fetal congenital heart disease (FCHD) remains a leading cause of infant mortality globally, yet the clinical deployment of deep learning models for automated fetal echocardiographic view recognition faces performance degradation due to domain shifts across medical centers. This study aims to develop a cross-center online adaptation framework to enhance model generalization in heterogeneous clinical environments without requiring target-domain labels. We propose a novel framework integrating three core components: (1) a dual-branch architecture with unadversarial perturbation-based consistency regularization to enforce feature invariance, (2) uncertainty-aware loss weighting via evidential deep learning (EDL) to prioritize high-uncertainty samples, and (3) selective fine-tuning of Batch Normalization (BN) layers to adapt domain-sensitive parameters efficiently. The framework dynamically adjusts model parameters during inference using unlabeled test data, avoiding costly retraining. Evaluated on multicenter fetal echocardiography datasets, the framework improved mean recognition accuracy by 0.88-2.55% across six deep learning models. DenseNet achieved the highest external test accuracy of 82.29%, outperforming baseline models. Feature visualization via t-SNE and heatmaps confirmed enhanced discriminative capability, while confusion matrices revealed reduced misclassification rates for challenging views (e.g., RVOT vs. 3VV/3VT). This work establishes a label-free adaptation strategy to address domain shifts in fetal ultrasound, demonstrating robust generalization across diverse clinical settings. By bridging AI innovation with practical deployment requirements, the framework offers a scalable solution to standardize prenatal screening quality, particularly benefiting resource-limited regions. Future efforts will extend validation to dynamic video analysis and broader multicenter datasets.

The purpose of this study is to develop and validate a multimodal, multitask prediction framework for clear cell renal cell carcinoma (ccRCC) by integrating preoperative CT radiomics,&#xa0;pathology-derived biomarker data from preoperative biopsy specimens, and clinical variables. The model was built for pathologically confirmed ccRCC and excluded other RCC histologic subtypes (e.g., papillary and chromophobe). In this multicenter retrospective study, ccRCC patients were enrolled and data were collected from preoperative CT scans, AMACR/P504S immunohistochemistry results, pathology reports, and follow-up records. Patients were&#xa0;assigned by hospital site&#xa0;into a&#xa0;training cohort&#xa0;and an&#xa0;independent external test cohort. Radiomic features were extracted from CT tumor regions of interest (ROIs), while deep learning features were derived from pathology images. Clinical variables were incorporated as additional inputs. A post-feature fusion strategy enabled simultaneous prediction of tumor classification and postoperative survival. Model performance was assessed using AUC for diagnosis and C-index for survival, together with calibration curves, decision curve analysis, and bootstrap confidence intervals. SHAP analysis was applied to quantify feature contributions. In external validation, the integrated multimodal model achieved strong diagnostic discrimination for P504S/AMACR (AUC&#x2009;=&#x2009;0.983) and demonstrated improved prognostic performance for postoperative outcomes (C-index&#x2009;=&#x2009;0.804). Subgroup analyses by grade and stage further supported model robustness, while SHAP-based interpretation indicated complementary contributions from imaging, pathology, and clinical variables. Overall, the proposed multimodal, multitask fusion framework enables reliable preoperative P504S/AMACR-based classification and postoperative prognostic prediction in ccRCC, supporting more refined risk stratification and individualized postoperative management using noninvasive imaging and clinical information. Calibration and decision curve analyses further support its potential clinical utility. Larger prospective and external multicenter validations are still needed to confirm generalizability.

Dual-energy subtraction (DES) radiography generates bone-suppressed (BS) and bone-enhanced (BE) images, enhancing soft tissue and bone visualization; however, it requires specialized dual-energy imaging systems. A deep generative translation model has been established to transform high-energy (HE) images into low-energy (LE) images (HE2LE), enabling virtual DES. However, clinical standard (STD) images cannot yield virtual DES via HE2LE translation alone. Therefore, we developed a U-Net-based STD into HE translation model using a discriminator in the pix2pix framework to generate virtual HE images and compute virtual DES from STD images. The training process involved a dataset of 600 triplet chest radiographs (STD, HE, and LE) and was executed over 2000 epochs using a sixfold cross-validation approach. The HE images generated by the STD2HE model were quantitatively evaluated and revealed a peak signal-to-noise ratio (PSNR) of 34.5, structural similarity index (SSIM) of 0.979, and deep image structure and texture similarity (DISTS) of 0.0257, compared with the groundtruth HE images. The virtual LE images obtained from virtual HE inputs were compared with the groundtruth LE images obtained by processing the corresponding groundtruth HE images using the same model. These LE images demonstrated a PSNR of 35.1, SSIM of 0.963, and DISTS of 0.0455. The virtual DES images exhibited a Fr&#xe9;chet inception distance of 67.6 and 73.6 for the BS and BE images, indicating high structural and perceptual fidelity, respectively. Integrating the proposed framework with the preceding HE2LE model can potentially enhance diagnostic utility in radiological analysis, emphasizing the need for specialized dual-energy imaging hardware.