Research On Cell Counting Based On Deep Learning

Cell counting is an important research content in medical image analysis,Accurate cell counting can detect potential diseases and related pathology.At present,various types of cells are still counted manually in clinic,but the detection efficiency of manual counting is low,and the accuracy will be affected by the status and experience of the inspectors.In recent years,due to the rapid development of deep learning,it has been widely used in object recognition,semantics segmentation etc.Deep learning technology applied in cell counting has also drawn many researchers'attention.Based on deep learning technology,the thesis studies the performance of different network models and architectures in cell counting through experiments.The main research contents are summarized as follows:1.Based on YOLO V3 detection framework,a low density cell counting method is proposed,which is mainly used for blood cell counting.According to the characteristics of blood cell data,a series of improvements are made to YOLO V3 algorithm:Using K-means clustering to calculate the anchor boxes size which is more suitable for detecting blood cells.A new multi-scale feature output method is adopted to improve the detection accuracy of small targets.Two residual units are deleted,the network is simplified and the detection speed is improved.The network output unit is replaced with Resnet unit,this avoids the disappearance of gradients and enhances the reuse of features.The non-maximum suppression algorithm is optimized by reducing the confidence score of the prediction boxes,to reduce the detection error for blocked red blood cells.Experimental results show that the proposed method has a significant improvement in the counting performance.Compared with other detection methods,the mean average precision(mAP)has been improved by more than 7%.2.Based on Fully Convolutional Networks and density estimation regression,a high density cell counting method is proposed,which is mainly used for bacterial cell counting.In order to improve the accuracy of counting,two cell counting networks A and B are designed.The input of the network can be any size of cell image,and the output is cell density map.Network A uses bilinear interpolation to upsample feature maps,and network B uses deconvolution structure to upsample feature maps.Experimental results show that network B achieves better counting performance.Its mean absolute error(MAE)is 2.5±0.2 and mean square error(MSE)is 2.9±0.1,which are lower than the current advanced high density cell counting methods.