Prospective assessment of breast cancer risk from multimodal multiview ultrasound images via clinically applicable deep learning.

ABSTRACT

The clinical application of breast ultrasound for the assessment of cancer risk and of deep learning for the classification of breast-ultrasound images has been hindered by inter-grader variability and high false positive rates and by deep-learning models that do not follow Breast Imaging Reporting and Data System (BI-RADS) standards, lack explainability features and have not been tested prospectively. Here, we show that an explainable deep-learning system trained on 10,815 multimodal breast-ultrasound images of 721 biopsy-confirmed lesions from 634 patients across two hospitals and prospectively tested on 912 additional images of 152 lesions from 141 patients predicts BI-RADS scores for breast cancer as accurately as experienced radiologists, with areas under the receiver operating curve of 0.922 (95% confidence interval (CI) = 0.868-0.959) for bimodal images and 0.955 (95% CI = 0.909-0.982) for multimodal images. Multimodal multiview breast-ultrasound images augmented with heatmaps for malignancy risk predicted via deep learning may facilitate the adoption of ultrasound imaging in screening mammography workflows.