Research On The Segmentation Algorithm Of Organs And Lesions In Medical CT Image Based On Deep Learning

Medical image segmentation is essential in medical image analysis.With the progress of modern medical treatment standard,the expeditious iteration of medical apparatus and the realization of universalization,high-level medical imaging technologies such as ultrasound,computed tomography（CT）,and magnetic resonance imaging（MRI）are widely used in various medical fields.Meanwhile,medical image segmentation plays an important role in medical image processing and analysis of blood vessels,skin lesions,tumor lesions and other human parts,provides a favorable basis for auxiliary medical diagnosis,and is of great significance to clinical treatment of real scenes.This article focuses on the research applied to real in clinical medical scenario of extraordinary potential value deep learning algorithm of medical CT image segmentation.Moreover,this article is committed to improving the accuracy of the algorithm and solving some defects of segmentation algorithm.The specific research results of this article are as follows:1、A multi-scale context-attention network for medical CT image segmentation（MC-Net）is proposed.Aiming at the problems of low segmentation accuracy of medical CT images,segmentation targets of different sizes cause learning difficulties,and large parameters of traditional models,this paper proposes a CT image segmentation network with multi-scale context attention.First,SP convolution（Split Convolution）is used to replace the traditional double convolution,which solves the problems of large parameters,general accuracy,and slow prediction segmentation in the classic U-Net.Then,the multi-scale feature extraction module is designed and added to the U-shaped backbone network composed of SP convolution,which improves the network’s segmentation ability to learn targets at different scales and segmentation accuracy.Meanwhile,after the down-sampling operation of the MC-Net segmentation network,a module based on atrous convolution is designed,it supplements the edge detail information lost due to continuous convolution layer and pooling layer operations.In addition,the residual attention module,the fast feature decoding module and the variant Dice loss equation are applied to the designed MC-Net,which further raises the prediction accuracy and inference time of medical CT image segmentation.The five presented modules and the backbone network are combined to form a large segmentation network.The experimental results show that the AC evaluation indexes on four different public medical CT image datasets are 0.9960,0.9946,0.9981 and 0.9924 respectively.The other three evaluation indexes are also better than the current advanced CT image segmentation network algorithm.2、A hierarchical context-attention transformer network for medical CT image segmentation（HT-Net）is proposed.In order to more effectively establish the long-distance dependence between segmentation target pixels and make up for the lack of self-attention ability in traditional convolution,this paper proposes a hierarchical context-attention transformer network for medical CT image segmentation.The network first established a residual hole pyramid pooling module,which learns segmentation targets of different scales by setting the hole pyramid convolution with different hole ratios,and also provides a wealth of position-sensitive axial attention modules that will be added later.The pixel detail information of the target edge.The position-sensitive axial attention module is designed based on the Transformer mechanism.The module uses the Transformer mechanism to calculate the similarity between the vertical and horizontal pixels of the input feature map to establish the self-attention between different pixels.Finally,this article also adds a three-input hierarchical context attention module to make up for the loss of the ability to capture local features caused by the Transformer mechanism.The three modules proposed above are sequentially added into the backbone network of Res Net-34,and used as a jump connection to assist the backbone network for target segmentation.In three 128 resolution public medical CT image datasets,the AC evaluation indexes of the algorithm are 0.9911,0.9924 and 0.9913 respectively.In the 512 resolution Lung and Ki TS19 datasets,the AC evaluation index is 0.9948 and 0.9944 respectively.Experimental shows that the segmentation ability of the network is better than the classic pure convolutional segmentation algorithm and the current popular Transformer-based hybrid Convolutional segmentation algorithm.