MRI-based deep learning and radiomics for severity classification of pediatric venous malformations

Venous malformations (VMs) are the most common congenital vascular malformations and exhibit high clinical heterogeneity. Current severity assessments heavily rely on subjective imaging interpretation and invasive angiography, which are limited by inconsistency and lack of standardization. Therefore, there is an urgent need for objective, quantitative, and automated tools to evaluate VM severity. We retrospectively enrolled 211 pediatric patients with peripheral VMs from a single center. Severity was defined by digital subtraction angiography (DSA) classification and dichotomized into localized/limited lesions (Types I–III) and diffuse/ectatic lesions (Type IV). All patients underwent routine MRI, and the proposed models were developed using T2 fat-suppressed (T2-FS) images. Radiomics features and deep learning–derived features were extracted and selected using the training cohort only and subsequently integrated with clinical variables to construct multiple classification models. The dataset was randomly divided into training and independent test cohorts. Performance was evaluated by ROC analysis, calibration, and decision curve analysis (DCA). Model interpretability was assessed using Grad-CAM. While univariate logistic regression revealed that coagulation parameters and lesion characteristics were associated with severity, only thrombin time remained significant in multivariate analysis. For imaging-based models, the deep learning–radiomics (DLR) model achieved an AUC of 0.990 in the training cohort and 0.957 in the independent test cohort. The combined model integrating clinical and imaging features achieved an AUC of 0.800 with an accuracy of 82.4% in the test cohort. DCA confirmed its superior net clinical benefit. We developed and internally validated an MRI-based deep learning and radiomics framework using routine T2-weighted fat-suppressed MRI for automated severity classification of venous malformations. The model demonstrated competitive and interpretable performance and may serve as a supportive decision-assistance approach pending further validation.