A multicenter deep learning framework integrating radiomics and vision transformers for comprehensive ovarian tumor analysis from ultrasound imaging

PURPOSE: This study aimed to develop and validate a robust multicenter deep learning pipeline that integrates radiomic descriptors with deep feature embeddings to enable comprehensive ovarian tumor analysis from ultrasound imaging, encompassing segmentation, multi-class classification, and prognostic prediction. METHODS: Ultrasound data from 3156 patients across eight centers were retrospectively analyzed. Five segmentation networks (UNETR, nnU-Net, Swin-UNet, SegNet, UNet) were trained to delineate tumors. From segmented regions, handcrafted radiomic features and deep features (ResNet, Vision Transformer (ViT)) were extracted. After reproducibility filtering (intraclass correlation coefficient (ICC) ≥ 0.75) and dimensionality reduction (PCA, RFE, ANOVA), three classifiers (TabTransformer, MLP, XGBoost) were trained for six-class categorization. Progression-free survival (PFS) was predicted using regression models. External validation was performed on 756 patients. RESULTS: UNETR achieved the best segmentation performance (DSC: 96.2%). For classification, the combined feature model with RFE and TabTransformer reached the highest accuracy (training AUC: 98.0%; external AUC: 95.8%; accuracy: 94.0%). For prognosis, TabTransformer achieved the best performance (C-index: 0.847), with consistent generalization across centers. Kaplan-Meier analysis confirmed significant survival group separation (p < 0.001). CONCLUSION: The proposed framework shows strong potential to reduce inter-operator variability and support personalized clinical decision-making in ovarian cancer care.