Medical Image Segmentation Algorithm Based On Feature Decoupling

Medical images reflect the structure of human organs and tissues,and their accurate segmentation has important reference value in quantitative pathological assessment and surgical guidance and prognosis.In recent years,deep learning-based methods,thanks to their powerful automatic feature extraction capabilities,have far surpassed traditional methods in terms of segmentation accuracy.However,the inherent difficulties of medical images,such as blurred lesion edges,foreign object occlusion and large lesion differences,pose great challenges to high-precision medical image segmentation.In order to further improve the segmentation accuracy,this dissertation proposes a multi-task segmentation framework based on feature decoupling to reduce the segmentation difficulty,and proposes two feature fusion strategies to construct a new network for the edge problem and the data distribution discrepancy problem,respectively.The main research contents of the dissertation are as follows:First,in order to reduce the overall difficulty of medical image segmentation,a multi-task segmentation framework based on feature decoupling strategy is proposed in this dissertation.The main idea of feature decoupling is to decouple the center and edge features from the segmentation features along the edge distance,so that these two relatively simple features can be extracted separately to reduce the difficulty of direct segmentation.In this dissertation,we use a two-branch multitasking framework to implement the decoupling strategy.First,we use center and edge labels to supervise the center and edge task branches respectively to obtain high-precision branching features,and then fuse the center and edge features again to obtain the final segmentation results.On this segmentation framework,two segmentation networks are built,and their excellent performance proves the effectiveness of the feature decoupling strategy.Second,in order to fuse the edge features and the center features effectively to get the final segmentation map with fine edges,this dissertation applies the 3D convolutional fusion strategy to the feature decoupling framework and proposes 3D-FDNet(3D convolution-based feature decoupling network).In this dissertation,the central features and edge features are regarded as the states of complete lesion features at different times,and then the two are combined into 3D features and fused using 3D convolutional computation in the time dimension.Compared with the normal linear fusion approach,this scheme benefits from the natural high aggregation capability of 3D convolution to enable a more adequate and complementary fusion of edge and center features under nonlinear computation.Experiments on the optic disc segmentation task show that the feature decoupling network based on the 3D convolutional fusion strategy in this dissertation far outperforms other advanced models in terms of segmentation accuracy.Third,in order to solve the challenge of network robustness caused by the small sample size and various lesion types of medical data,this dissertation applies dynamic convolution to the feature decoupling framework and proposes CCFDNet(conditionalsynergistic convolution-based feature decoupling network).In this dissertation,we design a dynamic convolution operator to generate a unique convolution kernel for each input to better fit the input distribution.Accordingly,this dissertation proposes a dynamic fusion module to dynamically fuse the central and edge features.In addition,to further improve the representational capability of the network,the backbone network is replaced with a Transformer architecture to achieve global dynamic performance.Experiments on polyp segmentation task and skin segmentation task demonstrate that the dynamic convolution-based feature decoupling network is more robust and greatly improves the segmentation accuracy compared with other standard convolution-based networks.