Left Ventricular Segmentation And Indicators Quantification Based On MRI

In the routine examination and diagnosis of clinical heart disease,the researches of left ventricular segmentation and indicators quantification have received great attention,which can evaluate the current situation of patients and then make appropriate and correct judgments on the diagnosis and treatment of the disease.In recent years,cardiac magnetic resonance imaging has become an important imaging method in the practice of clinical cardiac disease.Segmenting the left ventricle from cardiac magnetic resonance images refers to the division of the boundaries of the epicardium and endocardium.The indicators quantification of left ventricle mainly includes the prediction of two areas,six regional wall thicknesses and three cavity dimensions,which can provide reliable data support for the diagnosis and analysis of heart disease.Despite significant effort invested in left ventricular segmentation and quantification over the last decades,it is still under research,and due to the complexity of medical imaging,current algorithms are still not robust and flexible enough to support clinical practice.Therefore,the researches on left ventricular segmentation and indicators quantification still have great challenges,and there is still a lot of room for improvement in terms of robustness,flexibility and accuracy.Based on the above analysis,this paper designs and proposes a series of left ventricular segmentation and indicators quantification studies.The relevant studies are as follows:(1)Left ventricular segmentation based on multi-task constraints.According to the interrelationship between the left ventricular segmentation task and the regression task,this paper proposes a multi-task constraint network to automatically predict the segmentation result of the left ventricle,and uses the U-Net convolutional neural network architecture to segment the left ventricle as the benchmark.At the same time,considering the correlation between the tasks,we add multi-task constraints to the segmentation results,and use the inter-task correlation to constrain the loss function,which can improve the generalization of the network and the performance of the overall task.(2)Left ventricular segmentation based on Trans UNet and mult-itask constraints.Due to the lack of U-Net’s ability to capture long-term dependence,in this paper,we combine the Transformers and the U-Net to build Trans UNet deep network architecture to further improve the segmentation performance of U-Net.Trans UNet uses Transformers with self-attention mechanism as the encoder,which can better extract global information,combine with the decoder to extract local detail features,and add multi-task constraints for training,which further improves the performance of left ventricular segmentation task.(3)Comparison of left ventricular indicators quantification based on direct regression and segmentation.At present,the mainstream left ventricular quantification algorithms are mainly divided into two categories: direct regression based quantification methods and segmentation based quantification methods.This paper will build models based on these two different quantification methods and apply them to the indicators quantification of the left ventricle,including an end-to-end network for capturing cardiac structural features to estimate multi-type indicators for each cardiac MRI.And calculate estimation of the left ventricular indicators according to the obtained left ventricular MRI segmentation results.Two different quantification methods were compared and analyzed.In this paper,a series of related experiments are carried out under the same experimental conditions.The proposed and used related algorithms are evaluated and analyzed respectively,and compared with other mainstream algorithms.The experimental results show that the proposed methods have high accuracy in segmentation and quantification of the left ventricle.It is helpful to free doctors from tedious manual segmentation and quantification work,and has certain clinical significance.