Blood Cell Detection: Enhanced YOLOv9 Algorithm

Blood cell examination is a vital diagnostic tool in medicine, yet traditional methods suffer from low efficiency and accuracy due to overlapping cells, varying sizes, and high densities. To address these challenges, we present an enhanced version of the real-time object detection algorithm YOLOv9, incorporating a novel AKELAN4 module for improved blood cell recognition. Unlike previous approaches, AKELAN4 uniquely integrates Adaptive Kernel Convolution (AKConv) with Inverted Residual Multi-Branch Attention (IRMB), enabling simultaneous dynamic kernel shape adjustments and multi-scale feature extraction. This synergistic combination excels in detecting early-stage abnormalities and irregular shapes. To optimize model efficiency, we introduce the DySample sampling operator and Space-to-depth Convolution (SPDConv) module, reducing redundant calculations and enhancing feature learning efficiency. Our improved YOLOv9 algorithm achieved state-of-the-art performance on the Blood Cell Classification Dataset (BCCD), with 87.2% precision and 93.4% mean average precision, outperforming the recent YOLO-BC model by +6.3% and +3.4%, respectively. This work marks a significant advancement in blood cell detection research, with potential clinical applications.