Research On Images Segmentation Methods Of Pathological Slice Based On Deep Learning

The image segmentation of pathological slices is one of the most important steps in medical pathological image analysis.To some extent,the segmentation results may affect the accuracy and efficiency of pathological image analysis.The image segmentation methods based on deep learning have the characteristics of actively extracting features,which alleviates the embarrassing situation of the troublesome and rough segmentation results of traditional manual feature extraction.Therefore,it is of great significance and role to construct a deep learning-based medical pathological slice segmentation method in the field of medical images.Starting from the cell nucleus of medical pathological sections,and aiming at the existing problems and challenges in pathological images,this thesis constructs the segmentation algorithms of medical pathological sections.The main research works and conclusions of this thesis are as follows:(1)Aiming at the problems of small objects of interest,unbalanced position distribution,and indistinct boundary pixels in the task of nuclear segmentation,this thesis studies a segmentation algorithm based on coordinate attention residual network:DCA-Res UNet.Firstly,the algorithm constructs a depth separable residual module to extract more features.this algorithm uses the depth-separable residuals to extract more features.Secondly,a coordinate attention residual module is designed to focus on the remote distance of feature space,while highlighting the key information of nuclear location,and focusing on segmentation target pixels.Then the semantic information fusion module is designed to realize the fusion of image features at different levels,so as to improve the role of deep features in prediction and segmentation.Finally,in order to alleviate the image class imbalance,the algorithm combines the Dice loss function and the binary cross-entropy loss function for training.After experimental training on two public datasets DSB2018 and TNBC,the accuracy is 89.93% and 91.17%,respectively.(2)In order to further solve other problems existing in the algorithm in the previous chapter,this thesis constructs a segmentation algorithm combining hollow splitting attention: SAGU-Net,aiming at the problems of overlapping or adhesion of objects of interest,interference of background factors,and lacking of powerful feature interaction ability of existing segmentation algorithms in the task of nuclear segmentation.Firstly,the algorithm designs a cross-channel feature extraction module based on the integration of split attention and void convolution to achieve global and local multi-channel feature interaction.Secondly,the relevant information in the mixed space and channel dimension of the deep separable inverted bottleneck structure is designed,which can strengthen the attention of the weight of the segmentation target and improve the segmentation accuracy.Finally,the Ghost module is used to retain the details with similar features to make the model as lightweight as possible.After experimental training on public datasets DSB2018 and Mo Nu Seg,the accuracy of the algorithm is 92.86% and 92.51%,respectively.In addition,in order to verify the generalization performance of the algorithm,colorectal cancer pathological section images from legal cooperative units were self-generated into the dataset Colon Cancer in this thesis,and the accuracy of the algorithm was 90.15% after experimental training.The experimental results show that the pathological slice nuclear segmentation algorithm designed in this thesis have better segmentation performance both objectively and subjectively,and are better than other comparison algorithms.