A lightweight white blood cells detection network based on CenterNet and feature fusion modules

Abstract

White blood cells (WBCs) detection is significant to the diagnosis of many diseases. However, the detection accuracy can be influenced by the significant differences in color, size and morphology of WBCs in the images. In this paper, an improved CenterNet with smaller model size, fewer parameters, and lower computational complexity is proposed for the WBCs accurate detection. Firstly, the lightweight backbone GhostNetv2 is selected to reduce the model size, parameters and computational complexity of the network. Then, a feature pyramid network with the efficient channel attention (ECA) attention mechanism and the multi-scale feature extraction modules is constructed to enhance the capability of feature extraction and focus more on WBCs. Finally, the heatmap loss function is modified by proposing an improved mean squared error loss function to enhance the fitting ability between the predicted values and the ground truth of the heatmap. Experimental results show that the model size of the proposed lightweight CenterNet is only 19.9 MB, and the mAP.5 is 97.36%. The model size is reduced by 84% while the mAP.5 and FPS are increased by 0.7% and 10.4 compared to the original CenterNet. Moreover, the detection accuracy of the proposed lightweight CenterNet is comparable to the existing mainstream networks and its detection performance on different datasets is good, while the model size, parameters and computational complexity of the network is significantly reduced, and can be used for WBCs detection effectively.