EfficientNetV2-guided knowledge distillation for breast ultrasound diagnosis: Scalable deep learning with international clinical validation.

Abstract

e12571 Background: Breast cancer remains a leading cause of cancer-related morbidity worldwide, with imaging central to early diagnosis and treatment planning. Ultrasonography is widely used due to its safety and effectiveness in dense breast tissue; however, interpretation is operator dependent and varies across clinical settings. Although deep learning has advanced ultrasound-based cancer detection, many high-performing architectures impose substantial training and inference costs, limiting scalability. EfficientNetV2 introduces a training-aware design optimizing convergence and computational efficiency through progressive scaling and fused convolutions. We evaluated whether an EfficientNetV2-guided knowledge distillation framework could enable accurate, scalable breast ultrasound classification with international validation. Methods: We analyzed 8,116 breast ultrasound images (4,074 benign; 4,042 malignant) with pathology-confirmed diagnoses, independently reviewed by two board-certified radiologists with consensus adjudication. An EfficientNetV2 model served as the reference architecture. A compact EfficientNetB1 student model was trained using structured knowledge distillation with soft probabilistic supervision and regularized optimization. Performance was assessed on a held-out test cohort and externally validated on independent ultrasound datasets. Metrics included accuracy, sensitivity, specificity, F1 score, and AUROC. The distilled model was deployed within a cross-platform application and independently evaluated by over 40 physicians across six continents. Results: The EfficientNetV2 reference model demonstrated strong diagnostic performance ( > 99%). The distilled EfficientNetB1 achieved approximately 95% accuracy with balanced sensitivity and specificity and AUROC near 0.95 across external validation datasets. Performance remained consistent across heterogeneous ultrasound systems and geographic regions. More than 95% of the physicians reported the system to be clinically informative and suitable for routine diagnostic triage. Conclusions: Training-aware architectural design coupled with structured knowledge distillation enables lightweight deep learning models to retain clinically meaningful diagnostic performance while improving scalability and deployability. EfficientNetV2-guided distillation into an EfficientNetB1 architecture supports global implementation of AI-assisted breast ultrasound interpretation. Prospective studies are planned to assess impact on diagnostic confidence, biopsy utilization, and time-to-decision.