AE-YOLO: Feature Focus Enhancement for Breast Mass Detection

Abstract

Mammography remains the primary imaging modality for early breast-cancer screening. However, small mass size, irregular shape, and complex background tissue often limit the sensitivity and precision of computer-aided detection systems. In this work, we propose AE-YOLO, a novel enhancement of the YOLOv8 framework incorporating two key modules: aggregated dynamic convolution (ADC), which dynamically adapts convolutional weights across kernel, input-channel, and output-channel dimensions to strengthen feature extraction, and a visual enhancement block (VEB) comprising a lightweight transformer-based unit (TFormer) for global context capture and a feature reconstruction center (FRC) to suppress redundancy and refine mass features. Experiments on two public mammography datasets (DDSM and MIAS) demonstrate that AE-YOLO achieves a precision of 85.0%, recall of 77.2%, mAP50 of 84.9%, and mAP50:95 of 48.4%, outperforming current state-of-the-art models. Moreover, the proposed ADC and VEB modules are agnostic to network backbone and image source—they can be seamlessly integrated into other mammographic detection pipelines (e.g., INbreast) and consistently improve mass-detection performance across datasets and resolutions.