An Approach to Classifying Breast Density Level Using Deep Learning-Based Segmentation Model on Full-Field Digital Mammograms

Abstract

Breast density is a structure that determines the ratio of glandular and connective tissue in a woman's breast, directly affecting the assessment of breast cancer risk in radiological images. According to the Breast Imaging and Reporting Data System, it is classified on an international scale as follows: fatty (A), scattered density (B), heterogeneously dense (C), and extremely dense (D). The specified density categories can be examined in four views from a single patient, which are Right Craniocaudal, Right Mediolateral Oblique, Left Craniocaudal, and Left Mediolateral Oblique, using low-dose X-ray-based mammography. One of the major challenges in artificial intelligence-based breast density classification studies is the presence of the pectoral muscle, which affects the results in Right Mediolateral Oblique and Left Mediolateral Oblique views. Therefore, pre-processing methods are used to attempt automatic or semi-automatic removal of pectoral muscle from the images. This study proposes an approach for breast density classification using four-view mammograms from the VinDr-Mammo dataset, with the YOLOv8 segmentation module, while ensuring that the pectoral muscle is not excluded from the image. As a result of the study, successful breast density classification was achieved without removing the pectoral muscle, and the advantages and disadvantages of the system were discussed. © 2023 IEEE.