Pectoral Muscle Segmentation from Digital Mammograms Using a Transformative Approach
Dhanush Jain Mahaveera, Shubham Arun Gujar, Steven Cen, Xiaomeng Lei, Darryl Hwang, Bino Abel Varghese · 2023
Radiomics models based on digital mammography (DM) show the feasibility of reducing unnecessary invasive biopsies for benign mammographic masses while providing similar diagnostic performance for malignant mammographic masses as biopsies. Texture metrics, which comprises a majority of radiomics metrics assess the heterogeneity of radioopacities within the DM. The radiopacity of pectoral muscle (PM) and fibro-glandular tissue is comparable and hence difficult to separate from each other. The mediolateral-oblique (MLO) view of the DM is the most affected by PM. Subsequently, radiomic models constructed from MLO views of DM can hamper the accuracy of prediction models. Successful removal of pectoral muscle is challenging due to changes in shape, size, and texture of pectoral muscle in every MLO view DM. In this paper, our proposed algorithm is a technique of removing the pectoral muscle from MLO mammograms of the CBIS-DDSM dataset with a sample size of 800 images. By skillfully applying pixel clustering with thresholds, we precisely label clusters connected to neighboring pixels. This technique subsequently permits removing the cluster linked to the upper boundary of the image, yielding segmented mammograms, completely devoid of the pectoral muscle’s interference. Using Intersection over Union (IOU), an evaluation matrix widely used in computer-vision based object detection tasks.