A Comprehensive Two-Step Methodology for Automated Segmentation of Architectural Distortion in Mammograms

Breast cancer is one of the most prevalent forms of cancer worldwide and a leading cause of mortality among women. Early detection of breast cancer is crucial for effective treatment. Architectural distortion (AD) is an early sign of breast cancer, characterized by a subtle contraction of breast tissue that often goes unnoticed. Traditional methods of detection heavily rely on the expertise of radiologists, making the process more difficult. To address this, we propose a deep learning approach to automate precise and efficient AD segmentation. Our approach involves a two-step process. In the first step, we utilize a Mask R-CNN Detectron2 model to perform AD segmentation across the entire set of mammography images. This initial segmentation provides a baseline for identifying AD regions. The ResNet-18 patch model is incorporated into the Mask R-CNN model's segmentation pipeline in the second stage. The purpose of this combination is to improve AD area localization and precision. By combining the strengths of both models, we achieve improved accuracy in AD segmentation. The evaluation of our fully automated method yielded remarkable outcomes on a diverse test set consisting of private datasets, including the Baheya dataset and the NCI dataset, as well as the publicly available Digital Database for Screening Mammography (DDSM). The results showed a Segmentation Accuracy of 0.852, Classification Accuracy of 0.915, and Mean Average Precision (mAP) of 0.894. These findings demonstrate potential to enhance the efficiency and accuracy of AD detection and segmentation in mammogram images, contributing to early diagnosis and treatment planning for patients at risk of breast cancer.