Study On Medical Image Segmentation And 3D Reconstruction Technology Based On Deep Learning

Medical image segmentation is a prime technique to assist doctors in diagnosis with computer technology in modern medicine,and is of great significance to clinical disease monitoring,including lung segmentation,nucleus segmentation,multi-organ or tissue segmentation,etc.However,medical images suffer from low resolution,high complexity,large differences between different target regions,and lack of simple and effective linear features,leading to poor performance of models.Additionally,although 2D medical image segmentation is beneficial to clinical diagnosis,it fail to visualize human organs or tissues intuitively.The reason is that human tissues and organs are 3D and complex in reality.3D reconstruction of medical images aims to reconstruct medical images into organs or tissues to solve the problem,which can help doctors diagnose diseases more accurately.Medical image segmentation is the basis of 3D reconstruction,and the quality of image segmentation directly affects the application effect of 3D reconstruction in medical field.Aiming at addressing the issues existing in medical image segmentation and 3D reconstruction,this paper explores effective medical image segmentation strategies and3 D reconstruction methods.The main contributions of this paper are as follows:Firstly,in order to address the problem of complex segmentation targets and high demand for details in medical image segmentation,a medical image segmentation method with global and local information fusion is studied.And the proposed medical image segmentation architecture mainly consists of a global attention sub-network,local edge detection sub-network,feature encoding sub-network and feature decoding sub-network.Specifically,a global attention sub-network is designed to force the features extracted by the feature encoding sub-network to pay more attention to the target region and neglect the effect of irrelevant region.Then,the global overall features of the target region is obtained through the feature decoding sub-network.It is worth noting that,a local edge detection sub-network is designed to refine the edge details of the segmented target and improve the performance of the model,which uses the target contour information as additional supervision information.On the one hand,the contour information optimizes the ability of feature encoding sub-network to extract target features,and on the other hand,it can be combined with the overall target features output by the feature decoding sub-network to make the of results of models better.Secondly,for the problem of mutual interference between different target regions in multi-target medical images,a multi-target segmentation method based on multi-scale feature fusion is proposed.The proposed model mainly includes a feature extraction subnetwork,a multi-scale feature fusion sub-network and a feature decoding sub-network.Specifically,the model first learns the multi-scale features of segmented target in the image through a feature extraction sub-network.Then,a Transformer-based multi-scale feature fusion sub-network is designed to obtain the remote dependency between feature maps of different resolutions.In this way,the module can learn corresponding category features for different segmentation targets,so as to solve the problem of mutual influence and interference.Finally,we take advantage of fused multi-scale features in the feature decoding sub-network to achieve multi-object segmentation of medical images.Lastly,in view of the lack of spatial three-dimensionality of the 2D medical images,a3 D reconstruction method of medical images based on graph convolution is explored.And a framework which consists of a feature-shared sub-network,a fine-grained aggregation sub-network,a feature decoding sub-network and a deformation decoding sub-network is put forward in this paper.The framework firstly uses the feature-shared sub-network to encode sequence images into image sequence features.Then,the generated sequence features are input into fine-grained aggregation sub-network to generate aggregated edge contour features.At the same time,the feature decoding sub-network converts image sequence features into medical sequence segmentation images,and assists the generation of edge contours.Finally,the deformation decoding sub-network reconstructs the point cloud which is transformed from the image edges into a 3D mesh model based on graph convolution,obtaining the 3D form of the target organ.In a word,this paper proposes effective methods to solve the difficult problems in medical image segmentation and 3D reconstruction by studying the related key technologies,and these methods also have a certain value in practical applications.This paper also conducts sufficient experiments on multiple medical image segmentation and 3D reconstruction datasets,and the experimental results and corresponding analysis verify the effectiveness of the proposed methods.