Background parenchymal uptake classification using deep transfer learning on digital mammograms

Abstract

Background parenchymal uptake (BPU) in fibroglandular tissue on a molecular breast image (MBI) has been shown to be a strong risk factor for breast cancer and complementary to mammographic density. However, MBI is generally performed on women with dense breasts and only available at institutions with nuclear medicine capabilities, limiting the utility of this measure in routine breast screening and risk assessment. Digital mammography is used for routine breast screening. Our goal was to evaluate whether BPU features could be identified from digital mammograms (DMs) using deep transfer learning. Specifically, we identified a cohort of about 2000 women from a breast screening center who had DM and MBI performed at the same time period and trained models on DMs to classify BPU categories. We consider two types of classification problems in this work: a five-category classification of BPU and two combined classes. We designed and implemented machine learning algorithms leveraging state-of-the-art pre-trained deep neural networks, evaluated these algorithms on the collected data based using metrics such as accuracy, F1-score, and AUROC, and provided visual explanations using saliency mapping and gradient-weighted class activation mapping (GradCAM). Our results show that, among the experimented models, WideResNet-50 demonstrates the best performance on a hold-out test set with 58% accuracy, 0.82 micro-average AUROC and 0.72 macro-average AUROC on the five-category classification, while ResNet-18 comes out on top with 77% accuracy, 0.86 AUROC and 0.77 F1-score on the binary categorization. We also found that incorporating age, body mass index (BMI) and menopausal status improved classification of BPU compared to DM alone.