Abstract
Mammographic breast density has been established as an independent risk marker for developing breast cancer. Breast density assessment is a routine clinical need in breast cancer screening and current standard is using the Breast Imaging and Reporting Data System (BI-RADS) criteria including four qualitative categories (i.e., fatty, scattered density, heterogeneously dense, or extremely dense). In each mammogram examination, a breast is typically imaged with two different views, i.e., the mediolateral oblique (MLO) view and cranial caudal (CC) view. The BI-RADS-based breast density assessment is a qualitative process made by visual observation of both the MLO and CC views by radiologists, where there is a notable inter- and intra-reader variability. In order to maintain consistency and accuracy in BI-RADS-based breast density assessment, gaining understanding on radiologists' reading behaviors will be educational. In this study, we proposed to leverage the newly emerged deep learning approach to investigate how the MLO and CC view images of a mammogram examination may have been clinically used by radiologists in coming up with a BI-RADS density category. We implemented a convolutional neural network (CNN)-based deep learning model, aimed at distinguishing the breast density categories using a large (15,415 images) set of real-world clinical mammogram images. Our results showed that the classification of density categories (in terms of area under the receiver operating characteristic curve) using MLO view images is significantly higher than that using the CC view. This indicates that most likely it is the MLO view that the radiologists have predominately used to determine the breast density BI-RADS categories. Our study holds a potential to further interpret radiologists' reading characteristics, enhance personalized clinical training to radiologists, and ultimately reduce reader variations in breast density assessment.