SDE-YOLO: A Novel Method for Blood Cell Detection

Area of Science:

• Computer Vision and Image Analysis
• Biomedical Engineering
• Machine Learning for Healthcare

Background:

• Existing single-stage and two-stage algorithms exhibit limitations in blood cell detection, including low accuracy, misdetection, and missed detections.
• Accurate and efficient blood cell detection is crucial for clinical diagnostics and research.

Purpose of the Study:

• To develop an improved target detection algorithm, SDE-YOLO, based on the YOLOv5s framework.
• To enhance the accuracy and real-time performance of blood cell detection, specifically for white blood cells, red blood cells, and platelets.

Main Methods:

• Integration of Swin Transformer into the backbone for superior feature extraction.
• Modification of the path-aggregation network (PANet) by removing a network layer and employing depth-separable convolution to improve small target detection and reduce parameters.
• Replacement of the complete intersection over union (CIoU) loss function with the Euclidean intersection over union (EIoU) loss function to address sample imbalance and accelerate convergence.

Main Results:

• SDE-YOLO achieved high mean Average Precision (mAP) scores: 99.5% for white blood cells, 95.3% for red blood cells, and 93.3% for platelets on the BCCD dataset.
• Demonstrated superior performance compared to established algorithms like SSD, YOLOv4, and YOLOv5s.
• Outperformed YOLOv7 and YOLOv8 in terms of accuracy and real-time detection capabilities for blood cells.

Conclusions:

• The SDE-YOLO algorithm effectively addresses the challenges in blood cell detection, offering significant improvements in accuracy and efficiency.
• The proposed modifications, including Swin Transformer integration, PANet optimization, and EIoU loss, contribute to enhanced small target recognition and faster convergence.
• SDE-YOLO presents a promising solution for real-time, high-accuracy blood cell analysis in clinical and research settings.