Research On Medical Image Segmentation Based On Feature Encoding And Decoding Method

Medical image segmentation is one of the key techniques in medical image processing,which has important research value for assisting doctors in pathological analysis,diagnosis and treatment.Medical image segmentation is a process of dividing the image into many different parts according to some similar characteristics of the image,and then extracting the tissues of interest.However,as medical images are often interfered by various external factors in the imaging process,they are prone to problems such as serious noise,intensity non-uniformity and low resolution,thus losing a lot of valuable characteristic information,which brings great challenges to medical image segmentation technology.Although medical image segmentation methods based on deep learning emerge in endlessly,many segmentation networks do not sufficiently consider the characteristics of medical images,lack the utilization of multi-modality information,the extraction and fusion of hierarchical features,and the design of network robustness,so the segmentation effect is not ideal.Therefore,aiming at the shortcomings of existing segmentation methods,this paper proposes two novel segmentation methods to improve the accuracy and robustness of medical image segmentation.The main work of this paper includes:(1)A feature encoding and decoding method based on multi-modality information fusion is proposed.The proposed supervised methods do not sufficiently consider the characteristics of medical images and they are interfered by image noise and intensity non-uniformity,so the robustness and generalization of the methods are not good.Therefore,from the perspective of improving the feature encoding structure,firstly,based on the entropy rate super-pixel(ERS)segmentation algorithm,the multi-modality image is pre-segmented to obtain super-pixel blocks,and a new fusion algorithm is proposed to renumber them,then the super-pixel image is established.Then,each super-pixel block is represented by a node of the undirected image,and the feature vector is extracted by the gray value of each node.The correlation between nodes is judged by dissimilarity weight,and the feature sequence of adjacent nodes is constructed.Finally,the feature sequence is used as the input of bi-directional long short-term memory(BiLSTM)model.After training and testing,the final segmentation result is obtained.(2)A feature encoding and decoding method based on multi-scale feature fusion is proposed.Compared with the shortcomings of previous methods based on convolutional neural network that have not fully extracted the image feature information,the fusion and utilization of feature information of different scales are not sufficient,this paper considers the improvement of feature encoding and decoding structure,and designs a triple complementation network(TCN)that integrates transformer,convolutional neural network and recurrent neural network.This method makes transformer,convolutional neural network and recurrent neural network complement each other.In the encoder,transformer and convolutional neural network are used to extract and connect the feature maps of the input image to effectively capture the global and local feature information of the image.In the decoder,the recurrent decoding module(RDM)which contains recurrent neural network is used to refine the multi-scale feature maps obtained from the encoder,and the full-scale features are fused with dense connections to predict finer segmentation result.The experimental results of the two methods on the four medical datasets of BrainWeb,MRBrainS,BraTS2017 and Choledoch show that,the feature encoding and decoding method based on multi-modality feature fusion makes full use of the multi-modality information and spatial neighborhood information.Compared with the mainstream better algorithms,it greatly improves the precision of the boundary area division,and has a good anti-interference performance for local area noise and intensity non-uniformity.The feature encoding and decoding method based on multi-scale feature fusion makes up for the lack of the previous method which relies on the traditional manual methods to extract features.By combining transformer and convolutional neural network,it can extract richer feature information,and using the convolutional recurrent neural network to refine the multi-scale features,an end-to-end network structure is formed as a whole,which effectively improves the segmentation accuracy and robustness of the network.