Deep learning–based classification of benign and malignant breast lesions on ultrasound using knowledge distillation with external validation and global deployment.
Randomized trial developed a computationally efficient model for breast lesion classification, indicating its clinical utility across diverse settings.
Key Points
To create and validate an efficient deep learning model for classifying breast ultrasound lesions and assess its global clinical feasibility.
Analyzed 8,116 breast ultrasound images (4,074 benign; 4,042 malignant) annotated by radiologists.
Trained a high-capacity ResNet34 teacher network and a distilled ResNet18 student network using structured knowledge distillation.
Conducted external validation on three independent ultrasound datasets and evaluated the model's performance metrics.
Student model achieved 91.9% accuracy with sensitivity 90.8% and specificity 93.0%.
External validation showed stable generalization with accuracy between 89–92%.
94.2% of physicians in global evaluation found the system clinically useful for routine workflows.