Research On Organ Segmentation Algorithm Of Medical Image

With the promotion of medical imaging technology,it is increasingly common to use medical images to assist screening and diagnosis of diseases,which brings conve-nience to clinical medicine and also brings more and more onerous demands for film reading at the same time.Therefore,automatic film reading by computer has become a hot research topic recently.Organ segmentation is often a basic step in the review of organ images,and the precision of segmentation has a great impact on the subsequent screening and diagnosis operations,so accurate organ segmentation is of great signifi-cance.The research focus of this paper is to use the deep learning method to perform organ segmentation on medical images.Although a lot of recent works have been devoted to improving the accuracy of organ segmentation tasks and great progress has been made,organ segmentation accu-racy still faces some common problems.The problems concerned in this paper can be summarized into the following three aspects:1)the shape of organs varies greatly,and the lack of perception of the overall shape will lead to segmentation errors;2)There are great differences within the same organ category,and there are some regions of the same kind but with great differences in external manifestations,namely heterogeneous regions.Heterogeneous regions often lead to misdivision;3)The boundary of organs is blurred,and the pixels near the boundary are prone to segmentation errors.What's more,the boundary is often incomplete,and the labels of segmentation results often overflow in the segmentation results.In view of the above problems,this paper puts forward the following three solu-tions:1.In this paper,we firstly introduce a global context extraction module adapted to the shape of organs.Specifically,in order to increase the perception of the overall organ shape in the convolutional neural network,this module extracts the global context information of the feature map from two dimensions of space and channel respectively,and uses the global information to guide the offset learning of the deformable convo-lution.The offset learned in this way can obtain the convolutional sampling locations with the overall information of the organs,so as to adapt to the shape variation of the organs;2.In order to assist the segmentation of heterogeneous regions,this paper proposes a class-wise feature extraction module.Since the statistical characteristics of the het-erogeneous regions differ greatly from those of their categories but are correlated with those of pixels of the same category and the same image,this module extracts the cat-egory statistics information from the input image and fuses it with the network feature map to guide the fine-tuning of the segmentation results.The experiment proves that our module can correct the wrong predictions in heterogeneous regions.3.Some model enhancement algorithms for blurred boundaries are also proposed in this paper.Firstly,since the area near the boundary is prone to segmentation errors,this paper proposes a novel loss function,in which a higher loss weight is applied to the pixels near the boundary in later training period to correct the pixel-level segmentation errors near the boundary;In order to combat the interference of blurred boundary on the structure of segmentation results,this paper proposes a structural feature extraction module,which reserves more local structural information in the process of network feature extraction for optimizing the segmentation structure near the boundary.In this paper,the above methods are applied to two open-source organ segmenta-tion datasets.Based on the basic network,the accuracy of the pancreas segmentation dataset of the National Institutes of Health is increased by 9%,and the proposed methods receive accuracy improvement of about 1%and the best accuracy in the Japanese So-ciety of Radiological Technology chest radiograph segmentation dataset respectively.The application on these two datasets of different types,different imaging methods and different data dimensions not only shows the advantages of our method in performance but also demonstrates its generalization to different organ segmentation tasks.