Redesigned Dual-Task Learning Framework for Diagnosis Mammography Screening with BI-RADS and Density Classification

Mammography plays a pivotal role in breast cancer diagnosis and monitoring, yet the accuracy of Breast Imaging-Reporting and Data System (BI-RADS) assessments can vary among radiologists, particularly concerning breast density evaluations. Computer-aided diagnosis (CADx) systems have emerged to augment diagnostic precision. In this context, we propose a redesigned Dual-Task Learning (DTL) framework for mammography screening, focusing on BI-RADS and breast density classification. Our approach, notably DTL-Variant M, demonstrates superior performance across multiple metrics. DTL-Variant M showcases substantial enhancements in both BI-RADS and breast density classification tasks compared to other variants, empha-sizing its efficiency with ResNeXt-50 backbone. Furthermore, we employ focal loss, a highly effective loss function for imbalanced data, in our approach to tackle the problem of class imbalance and achieve better results for BI-RADS classification, which we consider more significant than density classification when using cross-entropy loss.