Research on deep learning-based blood cell detection method
Yao Liu, Peng Chen, Songyan Liu · 2024
An assessment of a patient's health status significantly relies on the analysis of blood components, particularly through the accurate enumeration of red blood cells, white blood cells, and platelets. Traditional algorithms for target detection typically encounter challenges such as inadequate accuracy, as well as frequent false positives and false negatives in the detection of blood cells. In this study, a multi-scale fusion blood cell detection model AGF-YOLOv8 is proposed, which firstly introduces the CBAM attention mechanism between the ConvModule and the Batch Normalization layer to improve the attention to key features; Secondly, it combines the structure of the SPPF and the FPN to realize multi-scale feature fusion, and adopts the Ghost module to reduce model parameters and computational complexity; Finally, to address sample imbalance, the Focal Loss is used in the Darknet Bottleneck to optimize the recognition of difficult-to-divide samples. Experimental findings indicate that, relative to the original YOLOv8, the detection accuracy of the AGF-YOLOv8 model has increased by four percentage points, which can provide auxiliary diagnosis for healthcare professionals in the future.