Research On Blood Cell Detection Algorithm Based On Deep Learning

Blood routine is a common examination in hospitals,which can play a greater role in clinical treatment.In blood routine examination,physicians often use manual microscopic examination and instrument counting method,which has low efficiency and accuracy.In recent years,deep learning has gradually become a popular research field.Compared with traditional methods,it can extract more image feature details and better image processing effects,so it has been widely used.Based on this,this paper proposes the method of using deep learning to detect blood cells,the specific research contents are as follows:(1)To address the problem of low detection accuracy caused by dense blood cells,mutual occlusion and varying cell size,an improved Centernet blood cell detection algorithm is proposed in this paper.In order to improve the accuracy of blood cell detection,this paper first introduces the ACON activation function in the backbone feature extraction network,which effectively improves the problem of poor blood cell detection accuracy when the network is very deep.Then,coordinate attention mechanism is introduced in the residual module of the backbone feature extraction network,which improves the problem of low detection accuracy caused by dense and mutual occlusion of blood cells.Finally,the ASPP module is introduced before feature fusion,which improves the problem of insufficient detection capability for blood cells of different sizes.The proposed algorithm was tested on the BCCD blood cell dataset,and the improved Centernet algorithm achieved a detection accuracy of 90.4% on the blood cell dataset.(2)To address the problem of unbalanced,dense and mutually occluded blood cell types in blood microscopy images that lead to the poor accuracy of existing blood cell detection methods,an improved YOLOX blood cell detection algorithm is proposed in this paper.The algorithm first introduces Focal Loss in the loss function to improve the imbalance of positive and negative samples and uneven cell types in the one-stage target detection algorithm.Then a hybrid attention mechanism is introduced in the residual module to reduce the probability of missed and wrong detection caused by mutual occlusion of blood cells.Then an adaptive spatial feature fusion module is introduced in the feature fusion tail to improve the feature expression capability.Finally,the inverse depth separable convolution module is introduced in the residual module to reduce the model parameters while also slightly improving the detection accuracy.The proposed algorithm is tested on the BCCD blood cell dataset and the improved YOLOX algorithm achieves 92.5% detection accuracy on the blood cell dataset.(3)To address the problem of low accuracy of blood cell detection due to the weak feature extraction ability of traditional convolutional neural networks,this paper proposes a blood cell detection algorithm based on Swin Transformer as the backbone feature extraction network.In this paper,the target detection algorithm Cascade RCNN is used as the base network,and the Swin Transformer network structure is used to replace the traditional convolutional neural network for feature extraction,thus enhancing the contextual information association and feature information transfer.Then,Bi FPN is used as the feature pyramid network of the algorithm for more effective fusion of information at different scales.The proposed algorithm was tested on the BCCD blood cell dataset,and the improved Cascade RCNN algorithm achieved 93.8% detection accuracy on the blood cell dataset.