BCG-YOLO: an efficient YOLO-based framework for multi-class blood cell detection in microscopic images

Automated blood cell detection from microscopic images is essential for routine hematological examination and computer-assisted laboratory diagnosis. However, accurate detection remains challenging because of the small size, dense distribution, overlap, and indistinct boundaries of blood cells. This study aimed to develop an efficient deep learning framework for accurate multi-class detection of blood cells in microscopic images.

We developed BCG-YOLO, an enhanced detector based on YOLO11n that integrates a lightweight backbone network, an improved multi-scale feature fusion module, and a dynamic detection head to improve feature representation and adaptive detection of small and densely distributed blood cells. The proposed model was evaluated on the public BCCD dataset and compared with representative one-stage detectors, transformer-based methods, and real-time object detection models. Ablation studies were conducted to assess the effectiveness of each proposed component. Additional five-fold cross-validation and paired image-level bootstrap analyses were performed for YOLO11n and BCG-YOLO.

On the original fixed test split, BCG-YOLO achieved an mAP@0.5 of 95.3%, an mAP@0.5:0.95 of 67.6%, and a recall of 89.8% while requiring only 3.80 GFLOPs. Under the common training and evaluation protocol used in this study, BCG-YOLO achieved the best overall detection performance among all evaluated models. Ablation experiments further demonstrated the effectiveness of the proposed modules, confirming consistent improvements in overall detection performance and computational efficiency.

BCG-YOLO provided a favorable balance between detection accuracy and computational cost in the internal evaluation on BCCD. These findings support its feasibility for microscopic blood-cell detection, while independent external validation is required to establish generalizability across laboratories, staining protocols, and imaging systems.