SKF-YOLO: An enhanced method for real-time blood cell detection

Abstract

In biomedicine, accurate detection of blood cells in microscopic images is essential for disease diagnosis. However, challenges like cell adhesion and overlapping often lead to missed detections and lower accuracy with traditional methods. To address these issues, this paper introduces an algorithm called SKF-YOLO, which builds on enhancements made to YOLOv11n. The algorithm incorporates several innovative components: a P6 detection head to improve the detection of large blood cells; the Single-Head Self-Attention (SHSA) module embedded in the backbone’s C3K2 module to enhance small-target localization in complex backgrounds; the KernelWarehouse module, which reduces the size of convolutional kernels while increasing their number for better computational efficiency; and the Focaler-MPDIoU loss function, derived from Focaler-IoU and MPDIoU, that emphasizes difficult samples to increase the model’s robustness. Tests on the BCCD blood cell dataset demonstrate SKF-YOLO’s superior performance, achieving a mean Average Precision (mAP) of 94.1 % and an Average Precision (AP) of 96.1 % for platelets. Compared to the baseline YOLOv11n, SKF-YOLO improves mAP by 2.6 % and reduces computation by 2.5 GFLOPs. These results confirm that SKF-YOLO outperforms other algorithms in blood cell detection and recognition, fulfilling the needs of lightweight target detection and offering valuable insights for future blood cell analysis in medical imaging.