Medical Image Segmentation Based On Multi-Scale Feature Fusion

Medical image segmentation is the process of dividing the areas of interest from medical images based on certain characteristics.It is a fundamental and indispensable step for quantitative analysis,3D reconstruction and registration.Beyond these fundamental aspects,medical image segmentation plays a critical role in clinical processes,including radiation therapy,image-guided surgery and pathological diagnosis.Its crucial role in these areas underscores the importance of continued research and development in the field of medical image segmentation.With the rapid development of deep learning,convolutional neural networks based on encoder-decoder structures are widely used in image segmentation.However,medical images are mostly single-channel grayscale images,with low contrast and low signal-to-noise ratio,and simple encoder-decoder networks cannot extract and retain fine-grained features and global context semantic features at the same time,resulting in difficulty in accurately segmenting medical targets with large shape differences.In addition,the damaged target is adhered to the surrounding tissue in the medical image,and the convolution is difficult to extract boundary features,resulting in the difficulty of completely segmenting the damaged tissue.In view of the above problems,this paper studies the multi-scale feature fusion strategy to improve the accuracy of medical image segmentation,and the main research work includes the following two aspects:(1)Existing encoder-decoder networks cannot capture and retain finegrained features and global context semantic features at the same time,making it difficult to accurately segment medical targets with large differences in shape.To solve this problem,a parallel multi-resolution encoder-decoder network is designed.First,a parallel multi-resolution encoder and a multi-resolution context encoder are designed.The encoder can extract and fuse multi-scale features from different encoding branches in parallel,effectively improving the feature representation capability of the network.The context encoder fuses the global context semantic features from different encoding branches to pinpoint the position information of the target to be segmented.Secondly,a parallel multiresolution decoder is designed.The decoder can continuously supplement the global context features of low-resolution branches to the feature maps of highresolution branches,effectively solving the problem of global context feature loss caused by up-sampling operations.In addition,the network is a common and flexible network framework that can meet the needs of different application scenarios by adjusting the number of network layers and the number of branches of parallel encoder-decoder.Experimental results show that the segmentation results of the network in four widely used medical image datasets are better than the current popular methods.(2)Due to the limited receptive field of convolution operation,it is difficult for the convolutional network to extract the boundary features of the target to be segmented and the surrounding tissue,resulting in the incomplete segmentation of the damaged tissue.In order to solve this problem,a multi-scale decoupled EM attention module is designed.Firstly,the design the decoupling EM attention branch,which decouples the attention weights into the sum of the pairwise iterm representing the inter-regional feature and the unary term representing the significance boundary feature under the premise of reducing the computational complexity,so as to extract the boundary features of the target to be segmented and the surrounding tissue.Secondly,a strategy for constructing multi-scale attention features is designed,so that the module extracts sparse and dense attention features at the same time.The sparse attention feature filters out noise interference and is suitable for segmenting small-sized targets;Dense attention features are suitable for capturing large-sized targets with more location information.In addition,the module is a plug-and-play module that can be applied to multiple medical segmentation tasks.A large number of experiments have been carried out on multiple public medical image datasets,and the experimental results show that the module can effectively improve the segmentation performance.