Research On Left Atrium And Scar Segmentation Based On Late Gadolinium Enhancement MRI

Atrial fibrillation is a common type of arrhythmia,which is related to the size of the left atrium and the changes in the myocardium.Cardiac magnetic resonance imaging(MRI)can measure the structure and function of the heart.However,common MRI technology cannot directly show the myocardial lesions,and late gadolinium enhancement MRI(LGE-MRI)technology is required to analyze and quantification myocardial lesions.To analyze and quantification of the size and scars of the left atrium,it is necessary to first identify and segment the left atrium and myocardial scars in the medical image.However,the drawback of LGE-MRI is that it can only enhance the contrast of scar tissue,but not clearly show the boundary between the left atrium and other tissues,which increase the difficulty and labeling cost of the left atrium.In addition,the distribution of myocardial scars is irregular,so it is necessary to first identify the location of the left atrial wall,and then set a threshold based on the gray difference of the scars around the left atrial wall to perform scar segmentation.This segmentation process is complex,and the accuracy of segmentation needs to be improved.Thereby,this thesis conducted the research on left atrium and scar segmentation on LGE-MRI.The main research of this thesis is as follows:(1)To solve the problem of blurred boundaries caused by the imaging characteristics of LGE-MRI,an uncertainty guided symmetric multi-level supervised left atrium segmentation method is proposed.In this method,we introduce the uncertainty information into the loss function to strengthen the model learning ability on these high-uncertainty regions while improving the boundary segmentation performance.The symmetric multilevel supervised network can calculate the uncertainty information of the model at one time without multiple times forward propagation.Moreover,the multi-level supervision network not only improves the detailed feature learning ability but also makes the distance between the middle layer and the prediction shorter,to make the model convergence faster.(2)To reduce the left atrium labeling cost on LGE-MRI data,a class-aware contrastive consistency semi-supervised segmentation method is proposed,thereby the model can learn with unlabeled data.This method constructs a contrastive consistency loss function on the class space to enable the class vector of unlabeled data to learn from the class vector of labeled data,allowing the model to extract more image features from unlabeled data while improving its discriminative ability.Moreover,this method introduces a rising strategy to dynamically control the weight of contrast consistency loss,and uses the class vector of labeled data as the reference for the unlabeled data class vector.Thus,the model can be trained more stably.The contrastive consistency loss function integrates the construction of the consistency regularity in semi-supervised learning and the construction of sample pairs in contrastive learning.It extends them to the category space,enabling this semi-supervised segmentation method to achieve perception capability of the categories.(3)To improve the segmentation performance of the model on small LGE-MRI training data,a few-shot left atrial segmentation method based on gray-level distribution information is proposed.This method is based on the medical image model in the MRI bias correction task.It represents the LGE-MRI as the product of the biased image and the real image,with the biased image serving as the characteristics of each image and the corrected real image representing the commonalities of the tissue.The proposed multi-task model combines the segmentation task with the grayscale bias correction task by designing a shared encoder for both tasks,the model can optimize them simultaneously and learn the characteristics and commonalities of images.Additionally,the proposed collaborative loss function guides the segmentation task using the real images,which allows the segmentation task to focus on the common features of the data,improving the model’s generalizability and robustness.Therefore,the model can segment test data better with a small number of training data.(4)To address the problem of complex segmentation process and low accuracy in scar automatic segmentation method,a joint segmentation method for left atrium and scar tissue is proposed by integrating soft region restrict and intensity difference constraint.This method explores the spatial and grayscale features of the left atrium and scar tissue.The proposed soft region restrict limits the scar prediction near the left atrium boundary to reduce the interference of other bright tissues.The proposed intensity difference constraint considers not only the gray difference between normal myocardium and scar tissues but also the intensity difference between the scar and the left atrial cavity.In addition,the design of joint segmentation can simultaneously predict the left atrium and scars without multi-stage processes,simplifying the process.This will significantly saves time for clinicians and improves clinical work efficiency.