Research On Segmentation And Classification Algorithm Of Thymoma CT Images Based On Deep Learning

Thymoma was a common thymic epithelial tumor that easily causes chest pain,myasthenia gravis,aplastic anemia,and other conditions,seriously affecting human health.With the rapid development of computer technology,computer-aided diagnosis of thymoma has become one of the current research hotspots.To effectively improve the accuracy of computer-aided diagnosis of thymoma,a series of improvements have been carried out in this study for the segmentation and classification of thymoma,as follows:Due to the complex tissue structure,blurred boundaries,irregular morphology,and different sizes of the lesions in the CT image of thymoma,and the lesions are usually connected to surrounding fat and blood vessels,resulting in low segmentation accuracy.From the perspective of multi scale feature extraction,a channel attention segmentation method based on Diverse Branch Block is proposed.Firstly,a multi branch feature extraction path is designed,and different scales of convolution are used to extract lesion features in the multi branch path to achieve effective extraction of different levels of thymoma features,in order to enrich the diversity of feature space;Then,channel attention mechanism is used in multi branch paths to improve the expression ability of features and further improve segmentation accuracy.From the perspective of detailed information such as the edge and contour of the lesion,an attention CGA segmentation method based on detail enhancement and content guidance is proposed.Different differential convolution algorithms are designed to calculate pixel values of detailed information such as edges and contours of lesions as prior knowledge to encode into CNN;Using the extracted features as content to guide the generation and full interaction of spatial and channel attention;An adaptive feature fusion mechanism based on CGA is designed to effectively fuse context features to improve segmentation accuracy.The advantages and disadvantages of Transformer and CNN in the modeling process are analyzed,and a thymoma CT image segmentation method based on Transformer and CNN is proposed,which extracts features globally and locally,respectively.Firstly,two parallel Transformer and CNN coding paths are constructed to extract global and local information of thymoma,respectively;Then,using the Bi Fusion fusion mechanism,the features extracted from the Transformer and CNN encoding paths are multi-level fused,and a progressive upsampling strategy is adopted to decode the multi-level fused features;Finally,the combined loss function is used for end-to-end training.Experimental results show that the results of this segmentation method are most similar to those of manual segmentation by clinicians.The problem of low classification accuracy for thymoma with different risk levels,The non IID characteristic of multicenter medical data leads to low generalization performance of classification models.a personalized federated learning classification method based on graph convolutional networks is proposed.First,in the model training process,mutual information is used to constrain the global model,client model,and personalized model for two-stage alignment training.Then,a personalized model is constructed using the client model parameters as node information using the adjacency matrix.Finally,an Encoder Decoder structure was designed to extract the focus features of thymoma CT images,and a personalized classification model was constructed based on the extracted features to classify thymoma with different risk levels.Experimental results show that this classification method has a high classification accuracy.