Early Breast Cancer Diagnosis via Breast Ultrasound and Deep Learning

Abstract

Low- and middle-income countries, such as the U.S.-Affiliated Pacific islands, suffer from much higher advanced stage breast cancer (Stages III and IV) rates than high-income countries, especially where mammography services do not exist or have low accessibility. Examples include Palau (77% of breast cancer cases diagnosed at an advanced stage), American Samoa (72%), and the Federated States of Micronesia (82%). Portable, handheld, AI-enabled breast ultrasound devices operated by a local healthcare worker could greatly reduce advanced stage cancer rates in the U.S.-Affiliated Pacific Islands by making screenings drastically more accessible. In this work, we have explored AI models for both breast lesion detection and breast density estimation from clinical breast ultrasound. Breast density assessment and lesion detection and diagnosis were trained and evaluated on task-specific datasets collected from clinical breast imaging centers across Hawaiʻi, available through the Hawaiʻi Pacific Island Mammography Registry. The results of the breast lesion detection task show that diagnosis of breast lesions is possible on ultrasound with concurrent classification of lesion descriptors for explainability, achieving 0.39 average precision. Precise delineation and classification of breast lesions is possible with AI applied to breast ultrasound. We expect performance to increase as more data become available. The typical performance across the breast lesion detection literature for non-explainable methods is 0.7 mean average precision. The breast density model is the first application of deep learning to predicting the BI-RADS mammographic breast density category from clinical breast ultrasound (inter-modality) and achieves 0.69 mean one-vs.-rest AUROC on a held-out test set. There is signal detectable by AI which relates mammographic breast density to breast ultrasound images. Methods for intra-modality classification of mammographic breast density with deep learning achieve approximately 0.93 mean one vs. rest AUROC on an internal test set.