Research on Efficient Detection of Blood Cells Based on YOLOv8 Integrated with CAFM and ODConv

This paper proposes an enhanced YOLOv8 model integrating the Convolution and Attention Fusion Module (CAFM) and Omni-Dimensional Dynamic Convolution Module (ODConv) to address the challenges of low detection accuracy and high computational complexity in blood cell detection. The CAFM module adaptively fuses multi-scale features through a hybrid convolution-attention mechanism, effectively capturing both global context and local details. Meanwhile, the ODConv module dynamically adjusts convolution kernels based on four-dimensional attention (spatial size, input channels, output channels, and kernel count), optimizing feature extraction for small and occluded cells. Extensive experiments on the BCCD dataset demonstrate that the improved YOLOv8 model demonstrated excellent scale adaptability. For smaller-scale datasets, the model could quickly converge during training and still maintain high detection accuracy, effectively reducing the risk of over-fitting. When dealing with large-scale datasets, the CAFM and ODConv modules could efficiently process massive amounts of information, showing strong generalization ability. The improved YOLOv8-N model achieves a mean average precision (mAP50) of 92.4%, surpassing the baseline YOLOv8-N by 3.3%. Specifically, the detection accuracy for white blood cells and platelets reaches 96.0% and 92.6%, respectively. Moreover, the model maintains lightweight characteristics with only 3.5M parameters and 7.7 GFLOPS, representing a 13.5% reduction in computational load compared to the baseline. Ablation studies validate the synergy between CAFM and ODConv, highlighting their complementary roles in enhancing detection performance. This work provides a promising solution for efficient and accurate blood cell analysis in clinical applications.