Boosting YOLO-based architecture for blood cell segmentation with ONNX acceleration and mosaic data augmentation

Accurate and efficient segmentation of blood cells in microscopic images is a fundamental task in hematological analysis, supporting early disease detection, longitudinal patient monitoring, and biomedical research. We present an end-to-end deep learning framework based on the YOLOv11n-seg architecture, designed for real-time instance segmentation and classification of blood cells. The model is trained and evaluated on the publicly available BCCD (Blood Cell Count and Detection) dataset, which contains annotated samples of red blood cells, white blood cells, and platelets. To improve the detection of small and densely overlapping cellular structures—particularly platelets mosaic data augmentation is employed during training, enhancing the diversity of spatial and contextual representations. The trained model is exported to ONNX and optimized using FP16 and INT8 precision, achieving lower latency and higher throughput than the ONNX FP32 baseline, which enables efficient real-time deployment on resource-constrained platforms. Quantitative evaluation shows strong segmentation performance, achieving an overall mask mean Average Precision (mAP@0.5) of 93.7%. Class-specific results further confirm the robustness of the approach, with mAP@0.5 scores of 92.8% for platelets, 89.3% for red blood cells, and 99.0% for white blood cells. These findings indicate that the proposed YOLOv11n-seg-based framework achieves clinically relevant accuracy while maintaining practical feasibility for real-time deployment.