Abstract
This study investigated a series of deep learning (DL) models for the objective assessment of four categories of mammographic breast density (e.g., fatty, scattered, heterogeneously dense, and extremely dense). A retrospective analysis was conducted using data collected from Taizhou Cancer Hospital over a period from January 2015 to December 2020. The dataset included mammograms from 9,621 women, totaling 57,282 images. The dataset was divided into training, validation, and test sets at a ratio of 7:2:1. Four DL models were employed, with Average Precision (AP) served as the primary evaluation metric. Additionally, the diagnostic performance of the DL models was compared with that of radiologists. Finally, we conducted validation of our model using an external test set. Among the DL models studied, InceptionV3 performed best, with AP values of 0.895 for almost entirely fatty, 0.857 for scattered fibroglandular tissue, 0.953 for heterogeneously dense, and 0.952 for extremely dense categories. The InceptionV3 model outperformed radiologists in accuracy and consistency. While radiologists surpassed the InceptionV3 model in fatty and scattered categories, their accuracy dropped significantly in heterogeneously and extremely dense categories. Nevertheless, our study demonstrated that DL can serve as a valuate tool in assisting radiologists with the objective quantification of breast density.