A YOLOv8x Framework for Breast Lesion Localization Using Real-World Mammographic Images

Breast cancer is one of the leading causes of mortality among women worldwide, and early and accurate mammographic assessment is important for timely diagnosis and improved patient survival. However, accurate breast lesion detection requires overcoming the issues of low lesion contrast, dense breast tissue, image noise, and variation in mammographic acquisition. In this study, we present a YOLOv8x-based computer-aided detection framework for automated breast lesion localization in full-field digital mammography (FFDM) and digital breast tomosynthesis (DBT) images acquired under heterogeneous clinical conditions. The framework was developed and evaluated using 501 clinical mammographic images obtained from multiple mammography systems, namely GE Digital Mammography for 2D FFDM and GE Senographe Pristina for DBT/3D mammography. The proposed approach integrated expert-guided radiological annotation, dedicated preprocessing, and high-resolution mammographic analysis at 1024 × 1024 pixels to improve lesion localization performance. Lesion regions were manually annotated by a specialist radiologist through clinically guided bounding-box procedures compatible with YOLO-based detection. Experimental results demonstrated accurate lesion localization performance, with precision of 98.87%, recall of 100%, mAP@0.5 of 99.50%, and mAP@0.5:0.95 of 99.16%. External validation further demonstrated promising cross-domain performance under heterogeneous imaging conditions. The findings demonstrate the potential value of clinically heterogeneous data, expert-guided annotation, and external validation for developing deployment-oriented computer-assisted breast lesion localization systems.