2026-09-29

Background: Breast MRI is the most sensitive imaging modality for breast cancer detection but requires time-intensive interpretation by experienced breast radiologists. Artificial intelligence (AI) has emerged as a promising tool to support image interpretation; however, robust external validation across heterogeneous clinical settings remains limited. Purpose: To develop a deep learning-based model for automated BI-RADS prediction from breast MRI and to externally evaluate its diagnostic performance on an independent dataset using a clinical reference standard for malignancy. Materials And Methods: In this retrospective multicenter study, a deep learning model was developed using 12,421 breast MRI examinations acquired between 2012 and 2024 at three institutions. The model was trained using radiologist-assigned BI-RADS categories as supervision targets and externally evaluated against a clinical reference standard for malignancy. Training data included examinations performed on 1.5-T and 3-T MRI systems using dynamic contrast-enhanced protocols. External validation was performed using an independent retrospective cohort of 158 breast MRI examinations acquired at an outpatient imaging center over a six-month period using different MRI hardware and imaging protocols. Diagnostic performance for malignancy was assessed using receiver operating characteristic (ROC) analysis and compared with expert radiologist assessment using the DeLong test. Results: The external validation cohort consisted of 158 examinations, including 21 malignant and 137 benign cases. Radiologist assessment achieved an area under the ROC curve (AUC) of 0.913 (95 % confidence interval [CI]: 0.857-0.952), whereas the AI model achieved an AUC of 0.767 (95 % CI: 0.693-0.830), corresponding to a statistically significant difference of 0.146 (P = 0.0001). At a BI-RADS threshold greater than 3, the AI model identified 54.7 % of all benign examinations as low-risk at a negative predictive value (NPV) of 97.4 %. Conclusion: The deep learning model showed lower discrimination than routine radiologist assessment but maintained high sensitivity at the evaluated threshold. Whether its identification of low-risk examinations can safely improve workflow requires prospective evaluation.

Objective: Incidental asymptomatic meningiomas are increasingly detected, yet predicting growth remains challenging. We aimed to determine whether a combined model integrating multiparametric MRI radiomics with clinical and semantic features improves growth risk stratification beyond conventional imaging in incidental asymptomatic meningiomas. Methods: This retrospective study included 519 patients from Institution A (training, n = 416; internal validation, n = 103) and 83 patients from two external institutions (external validation). Tumor growth was defined by volumetric criteria. Radiomic features were extracted from contrast-enhanced T1-weighted imaging (T1C), T2-weighted fluid-attenuated inversion recovery (T2-FLAIR), and apparent diffusion coefficient (ADC) sequences. A random survival forest (RSF) model combining all features was developed and validated. A simplified scoring system was derived from the most important features. Results: The combined RSF model demonstrated excellent discrimination, with C-indices of 0.901 (training), 0.837 (internal validation), and 0.813 (external validation), significantly outperforming a clinical-only model (ΔC-index up to 0.223). 9-point radiomics-based scoring system (cutoff = 5 points) stratified patients into high- and low-risk groups. The top predictive features included T1C (GLSZM GrayLevelNonUniformity, wavelet GLCM autocorrelation), T2-FLAIR and ADC texture features, and surface area. Conclusion: Multiparametric MRI radiomics improves growth risk stratification beyond conventional imaging. The simplified scoring system may support personalized surveillance while reducing unnecessary interventions, although prospective validation is required before clinical implementation.

Background: Ultrasound (US) is widely used for assessing lymph node (LN) status, but its diagnostic accuracy remains highly operator dependent. A robust computer-assisted diagnostic model may enhance clinical performance and improve inter-operator and inter-center consistencies. Purpose: To develop and validate a multimodal fusion model, ViT-Rad, that combines radiomics features and deep learning features derived from vision transformers (ViT) for the classification of benign and malignant LNs in US images. Methods: Between February 2016 and November 2023, a total of 1647 ultrasound images were retrospectively collected for analysis. In this multicenter study, we constructed ViT-Rad, a three-module neural network integrating ViT-based global contextual features and radiomics features extracted from manually delineated regions of interest. To address potential cross-center domain shift, we further employed weak/strong augmentation and a few-shot domain adaptation strategy using limited labeled external-center samples. Results: The model was trained and evaluated on a dataset from Center 1 (n = 1273; mean ± SD age, 57 ± 14 years), and its generalizability was tested on an external dataset from Center 2 (n = 374; mean ± SD age, 52 ± 18 years). ViT-Rad achieved an AUC of 0.95 [95% CI 0.91, 0.98] and an accuracy of 0.90 [95% CI 0.85, 0.95] on the internal test set, outperforming conventional radiomics models (AUC = 0.79, 95% CI 0.71, 0.89; P = .006). With domain adaptation, its AUC on the external set increased from 0.73 [95% CI 0.69, 0.79] to 0.85 [95% CI 0.81, 0.90]. These findings suggest improved adaptation-assisted external performance under cross-center domain shift. Conclusion: By combining radiomics and ViT-derived features, ViT-Rad effectively integrates domain-specific and global contextual information, improving internal diagnostic performance and showing improved external adaptability after few-shot domain adaptation for LN classification on US images.

Background: Cardiovascular disease (CVD) drives mortality in chronic kidney disease (CKD) patients. Conventional risk tools underperform in CKD, and few machine learning (ML) models undergo cross-national validation. We developed interpretable ML models for classifying CVD amongst Chinese and US middle-aged and elderly CKD patients. Methods: We conducted retrospective cross-sectional analyses using two nationally representative cohorts: NHANES (2011-2016, US) and CHARLS (Wave 1 and 3, China). A total of 1057 CKD patients from NHANES and 1109 from CHARLS were included. Seven ML algorithms were trained and optimized. Model performance was assessed on internal test sets and via bidirectional cross-population external validation. Model interpretability was achieved using SHAP (SHapley Additive exPlanations) analysis. Results: The XGBoost model performed best in NHANES (test AUC: 0.753), while the Logistic Regression model was optimal for CHARLS (test AUC: 0.769). During bidirectional external validation, both models showed a decline in discriminative performance when deployed across populations, indicating limited cross-population transportability. SHAP analysis revealed distinct feature contribution profiles: Age ranked as the top contributing feature in the US cohort, whereas hypertension was the primary contributing features in the Chinese cohort. Conclusion: CVD-associated feature profiles differ substantially between US and Chinese middle-aged and elderly CKD patients, supporting the necessity of population-calibrated classification models rather than applying a unified model across countries.

Purpose: The diagnosis of posttraumatic acute diffuse brain swelling (ADBS) poses a considerable clinical challenge. This study aimed to propose modified computed tomography criteria for accurately diagnosing posttraumatic ADBS and subsequently to develop and validate risk prediction models using machine learning (ML) algorithms. Materials And Methods: This retrospective multicenter study included 356 patients with traumatic brain injury from two medical centers, comprising 226 patients for model training and 130 for external validation. Posttraumatic ADBS was defined as the combination of diffuse injury type III and sulcal effacement at the vertex. Five ML algorithms were trained to develop prediction models for ADBS, with predictors identified using Least Absolute Shrinkage and Selection Operator regression and the Boruta algorithm. Model performance was evaluated using appropriate metrics to determine the optimal model, which was validated internally by cross-validation and externally in the validation cohort. The best-performing model was interpreted using SHapley Additive exPlanations (SHAP) and deployed as a simple risk calculator. Results: A total of 65 patients (18.3%) developed posttraumatic ADBS. Using four key predictors, the random forest (RF) model demonstrated excellent discriminative performance in both the derivation cohort (area under the receiver-operating-characteristic curve [AUC] = 0.973) and the external validation cohort (AUC = 0.957). The mean AUCs from five-fold and ten-fold internal cross-validation were 0.863 ± 0.042 and 0.859 ± 0.077, respectively. SHAP analysis provided global and local interpretability for the model, which has been deployed as a publicly available web-based risk calculator. Conclusions: The developed RF model shows potential for assessing the risk of posttraumatic ADBS, thereby contributing to clinical decision-making.

Introduction: This study aimed to develop and evaluate a ChatGPT-based application designed to assist multidisciplinary team (MDT) discussions in oncological urology. Patients And Methods: A ChatGPT-based tool (UroMDT Advisor) was developed to: (1) generate case summaries, (2) provide evidence-based treatment suggestions aligned with EAU, ESMO, and NCCN guidelines, (3) offer supporting references, (4) compare opinions across specialties to identify inconsistencies, (5) draft final MDT reports, and (6) create patient summaries. A consecutive series of anonymized oncological cases from eight centers were entered into the system. For each case, GPT-generated opinions were compared with those of MDT participants (urologists, oncologists, radiotherapists) and the final consensus decision. The model operated in a blinded fashion during concordance testing and was subsequently unblinded to produce final reports and patient summaries. Two expert urologists independently assessed the accuracy, completeness, and clarity of GPT outputs. Results: A total of 180 cases were analyzed: 123 (68.3%) prostate, 21 (11.7%) renal, 21 (11.7%) urothelial, 12 (6.7%) testicular, and 3 (1.7%) penile cancer cases. Exact agreement between UroMDT Advisor and urologists, oncologists, radiation oncologists, and final MDT decisions was 61.7%, 66.7%, 63.3%, and 66.7%, respectively; the corresponding unweighted Cohen's kappa values were 0.401, 0.489, 0.441, and 0.482. After resolution of discordant ratings, the GPT-generated final reports were rated as accurate in 90% of cases, fairly accurate in 8%, and inaccurate in 2%. Completeness was rated high in 95%, moderate in 2%, and low in 3% of cases, while clarity was rated high in 95% and moderate in 5%. Conclusions: The UroMDT Advisor showed moderate agreement with final MDT decisions and produced highly rated outputs. These findings support further evaluation of large language models as supervised decision-support and documentation tools in oncological MDT workflows while maintaining human oversight.