Comparative Analysis of YOLOv5 –YOLOv11 for Automated Detection of Breast Lesions in Mammographic Images

Abstract

Breast cancer is one of the most prevalent cancers in women worldwide, and mammography is the most important imaging technique to detect it at an early stage. Single-stage object detectors from the You Only Look Once family have gained more and more popularity in the field of automated breast lesion detection and classification in mammographic images in recent years. But a comparative study of all the generations of YOLOv5 to YOLOv11 has not been done for this task. This review consolidates 20 peer-reviewed studies and benchmark evaluations of the YOLO architectures ranging from YOLOv5 to YOLOv11 to examine and compare the detection accuracy, robustness across datasets, architectural innovation, and clinical readiness. The results suggest that YOLOv5 is the most advanced and widely tested family for mammography, achieving up to 88.0 mAP on INbreast and 84.3 after transfer learning from CBIS-DDSM. YOLOv8-based models show maximum flexibility such as micro calcification detection with mAP50 of 92.1. YOLOv9 made architectural improvements that were tested against CBIS-DDSM and YOLOv10 added NMS-free training. YOLOv11 is the best general purpose model, but its use in mammography is somewhat restricted. There is little literature on YOLOv6 in breast imaging. The main research gaps we have identified are: YOLOv10/YOLOv11 evaluations specifically for mammography; heterogeneity of datasets; and inconsistent reporting of metrics. Last but not least, we make suggestions for benchmarking.

Keywords: Breast Cancer, Mammography, YOLO, Object detection, Computer-Aided detection, Deep Learning