A pectoral muscle suppression approach for improved deep learning-based mammogram image analysis

Abstract

Breast cancer persists as a major health concern for women globally, and the best course of treatment depends on early detection. Although mammography is widely used as a monitoring tool, its limitations in accurately identifying subtle early-stage lesions and classifying malignant tumors persist. This research aims to develop an advanced mammogram analysis system that prioritizes the identification and classification of malignant tumors. The proposed methodology includes data preprocessing, pectoral muscle suppression, precise tumor localization, and subsequent classification into malignant or benign categories. To ensure a good level of precision in tumor detection, minimizing the disruption caused by the pectoral muscle is imperative. Effective suppression of muscle tissue improves image quality and facilitates precise identification of potential tumors. The publicly available CBIS-DDSM and VinDr-Mammo dataset were utilized for model training and testing. The proposed methodology, which integrates YOLOV8s with pectoral muscle suppression, achieved an accuracy of 97.94 ± 0 69%, a precision of 98.77%, and a recall of 96.98% when the CBIS-DDSM data set is used. An accuracy of 99.70 ± 0.15%, a precision of 100%, and a recall of 99.39% are achieved when using the VinDr-Mammo dataset. The combination of CBIS-DDSM and VinDr-Mammo is then used to train the model and is tested on a private dataset (NITC-MVR) to test its performance in a real-world clinical setting. This heterogeneous test resulted in an overall accuracy rate of 95.48% with a precision of 97.72% and a recall of 94.50%.