Deep learning analysis of hematoxylin and eosin-stained benign breast biopsies to predict future invasive breast cancer.

Abstract

Background: Benign breast disease (BBD) is an important risk factor for breast cancer (BC) development. In this study, we analyzed hematoxylin and eosin-stained whole slide images (WSIs) from diagnostic BBD biopsies using different deep learning (DL) approaches to predict those who subsequently developed breast cancer (cases) and those who did not (controls).

Methods: We randomly divided cases and controls from a nested case-control study of 946 women with BBD into training (331 cases, 331 controls) and test (142 cases, 142 controls) sets. We employed customized VGG-16 and AutoML models for image-only classification using WSIs; logistic regression for classification using only clinico-pathological characteristics; and a multimodal network combining WSIs and clinico-pathological characteristics for classification.

Results: Both image-only (area under the receiver operating characteristic curve, AUROCs of 0.83 (standard error, SE: 0.001) and 0.78 (SE: 0.001) for customized VGG-16 and AutoML, respectively)) and multimodal (AUROC of 0.89 (SE: 0.03)) networks had high discriminatory accuracy for BC. The clinico-pathological characteristics only model had the lowest AUROC of 0.54 (SE: 0.03). Additionally, compared to the customized VGG-16 which performed better than AutoML, the multimodal network had improved accuracy, 0.89 (SE: 0.03) vs 0.83 (SE: 0.02), sensitivity, 0.93 (SE: 0.04) vs 0.83 (SE: 0.003), and specificity, namely 0.86 (SE: 0.03) vs 0.84 (SE: 0.003).

Conclusion: This study opens promising avenues for BC risk assessment in women with benign breast disease. Integrating whole slide images and clinico-pathological characteristics through a multimodal approach significantly improved predictive model performance. Future research will explore DL techniques to understand BBD progression to invasive BC.