Research On Gland Cell Segmentation Method Based On Convolutional Neural Network

In recent years,adenocarcinoma has become one of the malignant tumors that seriously damage human health.The carcinogenesis of adenocarcinoma has a process of occurrence,development and invasion,but there are no obvious tissue characteristics in the early stage of carcinogenesis,so it is not easy to detect the adenocarcinoma in the medical examination.If we analyze the size,shape and other morphological manifestations of the gland cell image structure,it can be used as the basis for the early diagnosis of adenocarcinoma,and it is very helpful for the prevention and treatment of adenocarcinoma.The segmentation of glandular cell images is an important prerequisite for the analysis and interpretation of glandular cell images.Therefore,the accurate segmentation method of gland cell image has been proposed by more and more people,but there are still some problems.In this paper,based on the U-Net network,starting from the characteristics of glandular cell image,the difficult problems of glandular cell image segmentation are studied,and several different solutions are proposed.The main contributions are as follows:(1)An edge aware module is proposed,which is embedded in the U-NET model together with the atrous spatial pyramid pooling module,to solve the problems of unclear cell edge segmentation and information loss.This structure consists of two parts:a main codec stream for semantic segmentation and a shape stream for processing feature map at the boundary level.In the encoder part of the trunk,the convolution operation is constructed as residual block,and the atrous spatial pyramid pooling module is introduced in the last layer of feature extraction.In the decoder part of the trunk,the convolution operation is also constructed as the residual block,and the attention of the attachment is used to drive the decoder to highlight the edge information of the feature map extracted from the main encoder,and feed it to the trunk network to adjust the loss of model weight and the output of the trunk.The experimental results show that proposed network structure is better than U-Net,U-Net++,Psi-Net and ETNet,which are commonly used excellent medical image segmentation models.Compared with U-Net network,the Hausdorff distance of Warwick Qu and MoNuSeg data sets is decreased by 16.874 and 8.121 respectively,and the edge information is strengthened.(2)It is proposed that the convolution layer in the U-Net structure is combined into dense connection blocks.Without changing the number of existing network layers,it can learn cell features from different scales,extract image features to the greatest extent,and solve the problems of poor segmentation accuracy caused by different glandular cells in size,shape,and density.In addition,the improved network introduces the selfattention mechanism at the decoding part,which can establish a rich context-dependent model for local features.Thus,the remote dependence relationship can be modeled,unnecessary propagation can be suppressed,the accurate location of cells can be improved,and the accuracy of glandular cell segmentation can be improved.It solves the problem that the existing glandular cell automatic segmentation network does not make full use of the relationship between cells in the global field of view.The experimental results show that this network structure has excellent performance on the Warwick-Qu dataset of 2015MICCAI.Compared with the U-Net model,the F1 value increases by 2.4%,the Mean Dice increases by 2.55%,and the Hausdorff distance decreases by 2.7.(3)The generalization ability of the proposed models is verified.The proposed models are tested on several different cell data sets and digital retinal image DRIVE data sets.Compared with the U-Net model,the experimental results show that the network structure of this paper has better segmentation results,and for different data sets,the proposed model has its own applicability,which further shows that the generalization ability of the proposed model is better,and provides ideas for further expanding the scope of application of the model.