Clinic-aligned Dual Distillation of Video and Image Foundation Models for Automated Breast Cancer US Diagnosis

Purpose To evaluate the clinical applicability of the US Dual-Distillation model (USDist) through comparative analysis with state-of-the-art models, ablation analysis of dual-distillation components, and assessment on portable US devices. Materials and Methods This retrospective multicenter study evaluated USDist using US video datasets collected from 16 medical centers (August 2016-December 2024) and two independent public datasets. The model integrates spatiotemporal dual-distillation and dynamic-static feature fusion to transfer feature representations from video and image foundation models. Diagnostic performance was assessed using receiver operating characteristic (ROC) analysis with DeLong testing and Holm correction, along with evaluation of computational efficiency and qualitative feature visualization. Results A total of 5033 patients were analyzed (mean age, 49 years ± 12 SD; 5031 female). USDist achieved an average area under the ROC curve (AUC) of 0.95 (95% CI: 0.93-0.97) for breast cancer diagnosis across datasets. In the main cohort, USDist outperformed foundation models while using 98.3% fewer parameters. Across multicenter datasets, diagnostic performance was comparable to foundation models (all P < .05). On a portable US device, USDist maintained an AUC of 0.92 (95% CI: 0.86-0.95) with 4.1% of the computational cost of fullparameter fine-tuning. Conclusion USDist demonstrated high diagnostic performance for automated breast cancer diagnosis, with substantial parameter reduction compared with foundation models and maintained performance across multicenter and portable US settings. ©RSNA, 2026.