PERFORMANCE ASSESSMENT OF MALIGNANT MASSES DETECTION FROM DIAGNOSTIC MAMMOGRAPHY

Mass is one of the most common types of abnormalities in breast diseases. Identifying malignant masses in mammography prior to lesion biopsies remains a challenging task for radiologists. In this study, we aim to classify malignant and benign masses from diagnostic mammography in women who have been previously undergone ultrasound screening. In particular, we intentionally excluded lesions with well-established features that were readily identified by radiologists, and were already served as the basis for determining whether it's suspicious or not. This framework integrates segmentation and classification tasks using a two-step architecture to segment and classify those hard cases from 220 women (40 benign cases, 180 malignant cases) with 460 digital mammograms. Our selected model generated a cross-validated AUROC of 0.75 and a dice coefficient of 83.25%. This study demonstrates the potential of classifying benign and malignant masses from ultrasound-screened diagnostic mammograms, which could be used to explore additional clinical biomarkers and risk assessment standards that would help doctors identify suspicious lesions.