iSL-YOLOv11: global–local attention enhanced small lesion detection in mammography images

Breast cancer is one of the most common cancers among women worldwide, and early detection plays a crucial role in improving patient survival and treatment outcomes. In clinical mammography screening, however, detecting early breast lesions continues to be challenging because small lesions are often characterized by low contrast, blurred boundaries, limited visual saliency, and interference caused by overlapping dense breast tissue. These aspects make it difficult for detection models to accurately identify subtle lesion regions in complex mammography images. To address these problems, this paper proposes an improved YOLOv11-based network architecture, iAFF-SEAM-LAE YOLOv11 (iSL-YOLOv11), that combines global and local structure-aware feature enhancement for small-lesion detection in mammography images. Specifically, C3K2_iAFF is introduced to improve adaptive feature fusion across different semantic levels, SEAM is used to enhance local structural representation under tissue overlap and occlusion, and LAE is adopted to reduce redundant computation while preserving lesion-related texture information. Experimental results show that the proposed iSL-YOLOv11 model improves the mAP50-95 score from 0.677 to 0.784 compared to the original YOLOv11 baseline, achieving a relative improvement of 15.8%. The model also achieves an mAP50 of 0.935, showing a stronger localization ability for small lesions in mammography images. These results suggest that iSL-YOLOv11 can improve the detection of small and visually ambiguous breast lesions and may provide a potential technical basis for computer-aided mammographic lesion localization.