Automatic Breast Density Classification on Tomosynthesis Images

Abstract

Breast cancer (BC) is the type of cancer that most greatly affects women globally hence its early detection is essential to guarantee an effective treatment. Although digital mammography (DM) is the main method of BC detection, it has low sensitivity with about 30% of positive cases undetected due to the superimposition of breast tissue when crossed by the X-ray beam. Digital breast tomosynthesis (DBT) does not share this limi tation, allowing the visualization of individual breast slices due to its image acquisition system. Consecutively, DBT was the object of this study as a means of determining one of the main risk factors for BC: breast density (BD). This thesis was aimed at developing an algorithm that, taking advantage of the 3D nature of DBT images, automatically clas sifies them in terms of BD. Thus, a quantitative, objective and reproducible classification was obtained, which will contribute to ascertain the risk of BC. The algorithm was developed in MATLAB and later transferred to a user interface that was compiled into an executable application. Using 350 images from the VICTRE database for the first classification phase – group 1 (ACR1+ACR2) versus group 2 (ACR3+ACR4), the highest AUC value of 0,9797 was obtained. In the classification within groups 1 and 2, the AUC obtained was 0,7461 and 0,6736, respectively. The algorithm attained an accuracy of 82% for these images. Sixteen exams provided by Hospital da Luz were also evaluated, with an overall accuracy of 62,5%. Therefore, a user-friendly and intuitive application was created that prioritizes the use of DBT as a diagnostic method and allows an objective classification of BD. This study is a first step towards preparing medical institutions for the compulsoriness of assessing BD, at a time when BC is still a very present pathology that shortens the lives of thousands of people.