Left Ventricle Segmentation And Quantification Based On Semi-supervision And Signed Distance Function

Cardiovascular disease is the world’s leading cause of death.Early detection and treatment can help to reduce patient mortality.In the diagnosis of cardiovascular diseases,left ventricle segmentation and indicators quantification on magnetic resonance images of the heart is an essential procedure for assessing the severity of patients with heart failure,This method,which is noninvasive and efficient,and widely used in clinical diagnosis of cardiovascular diseases.More specifically,left ventricle segmentation refers to the accurate delineation of the left ventricle on the cardiac images and the estimation of metrics such as the area of the segmented region to quantify the left ventricle ejection fraction and provide an important numerical reference for the diagnosis and classification of heart failure.Currently,the main challenges of fully automated research in left ventricle segmentation and quantification are: 1)the complexity of medical imaging leads to difficulties in feature extraction for left ventricle segmentation and poor segmentation results;2)difficulty in obtaining a large number of labeled left ventricle images and poor migration of supervised learning-based segmentation algorithms in practical applications;3)lack of accurate indicators quantification algorithms.This paper investigates the fully automated left ventricle segmentation and indicators quantization methods based on magnetic resonance images to address the above challenges.The main contents of this paper are:1)A left ventricle segmentation algorithm based on UNet 3+ and Transformer is proposed,aiming to better extract the image features of the left ventricle.The algorithm captures both fine-grained and coarse-grained image semantics using UNet 3+ and extracts global information of cardiac MRI images using Transformer to achieve automatic segmentation of the left ventricle.Also,a shape-aware representation is provided for the segmentation by jointly predicting the segmentation probability map and the signed distance map to further improve the performance of the left ventricle segmentation task.2)To address the problem of sparse labeled data,a semi-supervised left ventricle segmentation algorithm is proposed in this paper and trained in conjunction with an adversarial network.The algorithm uses U-Net as the segmentation network and treats it as a generator in the generative adversarial network architecture.Meanwhile,a discriminator based on a full convolutional structure is designed.Its output is used as a supervised signal to discriminate whether the signed distance map output from U-Net is tagged or not,and the adversarial loss is introduced in the training phase to implement the semi-supervised left ventricle segmentation algorithm based on the adversarial network.By this method,it is possible to make full use of a large amount of unlabeled data to improve the performance of the segmentation algorithm.3)To address the problem of insufficient accuracy of left ventricle indicators quantification algorithms,this paper proposes segmentation-based and regression-based methods to quantify the left ventricle.Among them,the segmentation-based approach calculates the left ventricle ejection fraction based on the segmentation results of the left ventricle,while the regression-based approach combines CNN and LSTM to estimate each index of the left ventricle and use it to calculate the ejection fraction.In summary,the research in this paper provides a series of effective algorithms for left ventricle segmentation and indicators quantification and compares them with several existing mainstream algorithms.The experimental results show that the proposed method in this paper has high accuracy in both left ventricle segmentation and quantification,and reduces the training data required for experiments,which is beneficial for physicians to facilitate the diagnosis of cardiovascular diseases.