Research On Cross-modal Cardiac Image Domain Adaptive Segmentation Algorithm

Cardiovascular disease is one of the most common diseases leading to human death in the world.Prevent and effectively treat cardiovascular disease is a hot topic in global research.The heart is the most important organ in the cardiovascular system.Through the study of each substructure of the heart,the performance indexes and pathological characteristics of the heart can be obtained.Therefore,the accurate segmentation of the heart site plays a very important role in the prevention and treatment of cardiovascular diseases.In practice,the diagnosis of heart diseases often adopts a variety of imaging methods for comprehensive analysis.It is of great clinical significance to use images of different modes to carry out the same analysis task.In addition,each substructure of the heart has overlapping or fuzzy boundary,which makes the task of heart segmentation more difficult.Aiming at each heart substructure in cardiac segmentation task context information contact,segmentation,the problem of low accuracy and efficiency in order to fully tap and use different modal correlation characteristics of the heart of the image,this topic proposed a cross modal antagonism domain adaptive full heart segmentation algorithm,the concrete work is as follows:(1)Based on Generative adversarial network,cross-modal medical image domain adaptation was realized.The traditional Unet model was integrated,and the trunk feature extraction coding network was redesigned.Combined with Resblock module,a feature extraction module was added by Dalited Convolution.The feature extraction performance of the model is improved by deepening the feature extraction network and increasing the receptive field.(2)To solve the problem of low segmentation accuracy when the heart pixel proportion is relatively small,multi-attention mechanism of space and channel is introduced,and an improved multi-attention module(MASC)is proposed.By using the mask weighting of the attention layer,the feature extraction network pays more attention to the key features of some feature layers and the spatial region of the heart image,while suppressing the invalid features of the spatial region such as the background of the heart image.(3)The cross-modal image adaptive algorithm and the improved segmentation algorithm are fused,and the algorithm shares the network parameters in the feature extraction stage,so that the adaptive and segmentation tasks can benefit each other.Finally,the algorithm realizes the adaptive transformation of CT and MR modes at image level and feature level,and also realizes the end-to-end model training and testing for adaptive tasks and segmentation tasks.Finally,the method proposed in this paper was trained and tested on the datasets of the whole heart segmentation and glioma segmentation competitions provided by the MACCAI Society.Among them,the cross-domain adaptive image conversion experiment uses the whole heart to segment the two-dimensional slice data,and compared with the results obtained by the existing algorithms,a better visual result is obtained.Before the heart segmentation experiment,the glioma data was used as the auxiliary experiment to test the performance of the algorithm,and then the heart segmentation data set was used to test the cross-modal data segmentation of the fused algorithm,and the more accurate segmentation results were obtained.