Research On Gland Cell Segmentation Algorithm Based On Deep Learning

Rich morphological statistics are an important basis for early diagnosis of adenocarcinoma,and accurate gland segmentation is a key prerequisite for obtaining reliable morphological statistics.Therefore,gland segmentation plays an important role in medical image processing.The development trend of medical image processing indicates that the segmentation model based on deep learning is an important method of gland segmentation.At present,there are three difficulties in the application of deep convolutional networks to gland image segmentation,including the problem of decreasing feature resolution,the existence of multiple scales of gland cells to be segmented,and the inherent invariance of deep convolutional networks.In view of the above three problems,this paper proposes two segmentation methods based on the improved U-Net model.The first segmentation method is a non-end-to-end improvement method that combines a traditional U-Net model with a fully connected conditional random field.The second segmentation method is an end-to-end improvement method that changes the internal structure of the traditional U-Net model.The non-end-to-end improvement method proposed in this paper refines the taiget edge by directly applying the fully connected condition random field to the segmentation graph of the traditional U-Net model output;the end-to-end improvement method adds the cavity residual module and the cavity.The Pyramid Pool Module and Attention Module improve the three major issues mentioned above.In order to make the two improved models have better performance,the parameters of the model are selected through multiple test experiments,including the parameters of the fully connected conditional random field and the void ratio of the cavity convolutional layer.Finally,the segmentation performance of the two improved models was verified based on the Warwick-Qu dataset,UCSB breast cancer cell dataset and MoNuSeg dataset.The generalization performance of the two improved models was verified based on the PatchCamelyon dataset and the DIC-HeLa dataset.The experimental results show that the segmentation performance of the non-end-to-end segmentation method of this topic has no outstanding advantages compared with other excellent models,while the end-to-end segmentation method shows more detailed segmentation results than other models.