Research On Medical Image Segmentation Methods Based On Encoder And Decoder

Medical image analysis is an important part of clinical medical diagnosis,and accurate medical image segmentation methods play a crucial role in assisting the analysis of medical images.Instead of relying on manual annotation of segmentation targets in the past,it greatly reduces labor costs,makes medical image segmentation more automatic and intelligent,and provides reliable basis for disease diagnosis.There are several problems in medical image segmentation tasks: it is difficult to identify small targets;The ability to extract boundary information is weak;The scale of segmentation targets such as lesions or organs varies;The demand for real-time application to mobile terminals or Edge device is high.The main research content of this article is as follows:A segmentation method called Strong UNet is proposed to address the difficulties of small object segmentation and weak edge information in medical images,which combines shifted window self-attention and convolution.Using convolution to extract shallow features,small receptive field help to obtain fine grained information;Simultaneously using the window local self-attention mechanism to enhance and extract local information.Modeling global information through shifted window self-attention mechanism helps to solve the problem of difficult segmentation between similar targets or similar targets and the background.In order to ensure that the accuracy of the method is basically unchanged or increased when deployed on mobile terminals or edge devices,and to reduce the number of parameters as much as possible,a plug and play lightweight convolution and cross attention block are proposed,and a multi-scale depth separable convolution with a inverted residual structure is constructed to extract local features.A structure similar to standard convolution is designed to be expanding,cross attention global processing and tucking to enhance the ability of spatial inductive bias and extract global information.In response to the high complexity of the shallow network module in the first method,a lightweight medical image segmentation method,Lighter Net,is proposed to further deepen the direction of lightweight based on this method.The more complex modules in the network are replaced by plug and play lightweight convolutions and cross attention blocks.The method uses block fusion for downsampling to reduce information loss.At the same time,we use skip connection and concatenation to fuse high-level and low-level Semantic information,and achieve accurate segmentation of multi-scale fusion.In summary,this article proposes a medical image segmentation method called Strong UNet based on shifted window self-attention and convolution fusion,as well as a segmentation network called Lighter Net based on lightweight convolution and cross attention blocks.The performance of the method has been experimentally verified on public datasets,and compared with other methods,it has better segmentation performance.