MMC-YOLO: Application of YOLOv8 in the Detection of Small Lesions in Biomedical Images

In the domain ofbiomedical image analysis, object detection is pivotal for the identification of lesions. Despite advancements in lesion recognition and localization, current methodologies exhibit limitations in accurately detecting minute biological entities. To address this challenge, this study presents an innovative medical object detection approach based on the YOLOv8 architecture, termed MMCYOLO. This approach leverages the advantages of Convolutional Neural Networks (CNNs)[1] and transformers by developing an Integration Module with Hybrid Attention (IMHA), which integrates convolutional layers with a multi-head self-attention mechanism[2], and a Multi-scale Feedforward Network (MSFN)[3]. These components address the limitations of convolutional kernels in long-range information interactions and transformer feedforward networks in single-scale feature aggregation. The multi-head self-attention mechanism autonomouslyfocuses on disease-relevant regions, while the MSFN enhances multi-scale information aggregation. Collectively, these dual mechanisms substantially improve the model's performance. Experiments conducted on the BCCD dataset validated the eficacy of MMC-YOLO, demonstrating its superior capability in detecting and localizing complex small lesions compared to existing methods.