Research On Medical Image Segmentation Based On Encoder-Decoder Structure

Computer-aided diagnostic systems that employ medical image segmentation technology play a critical role in clinical applications.The primary goal of medical image segmentation is to separate specific regions from an image to assist physicians in accurate disease diagnosis,assessing the severity of the disease,formulating treatment plans,and tracking disease progression.However,achieving high-precision segmentation remains a challenging task due to the complex characteristics of medical segmentation tasks,such as multi-scale,blurry boundaries,and multiobjective.When performing multi-scale object segmentation tasks,there often exist small missed target areas and difficulties in dealing with irregular,curved,and narrowed shapes.In addition,many medical image targets have blurry boundaries and low contrast with surrounding normal areas,and there may be noise interference in the background area,all of which can affect the accuracy of target segmentation.In particular,for multi-organ segmentation tasks,multiple targets need to be simultaneously processed to cope with the mutual correlation and interference between organs.This paper focuses on three typical tasks in the complex medical image segmentation mentioned above.To address the shortcomings of existing models,corresponding medical image segmentation methods are proposed based on the encoder-decoder architecture to achieve highquality segmentation targets of medical images in different scenarios.1)This paper proposes a dual-branch decoder network based on multi-scale context perception for multi-scale irregular segmentation tasks.Firstly,through the context information extraction module,the receptive field of the model is enhanced to improve its ability to extract targets of different scales.Secondly,by introducing the auxiliary branch decoder module and using decomposed convolution to extract feature information of irregular regions.Finally,the abstract information contained in the deep stage is fully integrated to alleviate the loss of semantic information during the upsampling and restoration process of the main branch decoder,and to solve the problem of small targets being lost during upsampling.Through quantitative analysis,qualitative analysis,and ablation experiments on four publicly available datasets(CVC-Clinic DB,Kvasir,Gla S,Data Science Bowl 2018),the effectiveness of the model is demonstrated.2)This paper proposes a cascaded encoder-decoder segmentation network to address the issues of low contrast,noisy background,and blurry boundaries in medical image segmentation tasks.Firstly,the first encoder-decoder network generates an initial mask from the input image,and then the pixel-wise multiplication of the mask and input image enhances the contrast between the object and background and eliminates the background noise.Secondly,basic convolution blocks based on depthwise separable convolution are used to reduce the model’s parameters and computation.Finally,a lightweight attention module is proposed,which uses non-local attention mechanism to improve the network’s understanding of global contextual information,making it easier for the model to obtain important features related to blurry boundaries.Through comparative experiments and visual analysis on the blurry edge ISIC dataset,the results show that the proposed method can effectively address the aforementioned issues.3)This paper proposes a CNN-Transformer hybrid encoder model that integrates convolutional neural networks(CNNs)and Transformer modules to address the issues of mutual correlation and interference in multi-object segmentation tasks.First,the hybrid encoder can simultaneously obtain high-detail local information and establish long-range dependencies,complementing each other’s advantages and improving the model’s performance in multi-object segmentation tasks.Second,a convolutional block attention module is used to reduce redundant feature information,making the extracted features more focused.Finally,a feature fusion module is adopted during the skip-connection process to perform multi-stage feature fusion,fully integrating shallow spatial information with deep abstract information.The proposed approach is validated on the Synapse multi-organ dataset through comparative experiments and ablation studies,and the results demonstrate the feasibility of the proposed method for solving multi-object segmentation problems.