Deep Learning Model And Algorithm Research For Heart Segmentation Based On Multiple Image Data

A variety of cardiovascular diseases are closely related to ventricular function,and the calculation of ejection fraction plays a very important role in assessing ventricular function.Accurate calculation of the ejection fraction depends on the annotation of the ventricular region,which is time-consuming and subjective for clinicians to annotate manually.Therefore,it is important to use a computer-assisted approach to automatically segment the ventricular region,which can save the physician’s time in annotation and make the calculation more objective.MRI and 2D ultrasound are the two most commonly used cardiac examination modalities.The irregular shape variation of the right ventricle in cardiac MRI and the small target of slices near the apex are difficult to segment,and it is even more difficult to achieve automatic segmentation in echocardiography due to unclear imaging and lack of edge information.In this paper,we realize automatic segmentation of the heart based on MRI images and 2D echocardiography,and the main research results are as follows:In this paper,the right ventricular segmentation study is carried out using the RVSC public dataset,the training data is expanded using data augmentation techniques,the U-Net model is used for experiments on the RVSC dataset,the residual module,the multi-scale dilated convolution module,etc.are introduced to optimize the network structure,and the network training is divided into two steps of localization and segmentation to effectively solve the problem that the target area near the apical region is too small.For the algorithm,the Tversky loss function is improved for the small target segmentation problem,and the segmentation accuracy of slices near the apices is improved.The results of a series of ablation experiments such as comparison experiments with advanced models in this field,replacement of the feature extraction network of the improved model,and introduction of new structures into the new model show the better performance of the improved model.The ejection fraction is calculated based on the segmentation results of the model,and the obtained ejection fraction is analyzed for consistency with the physician labeling results,and the high correlation coefficient indicates that the improved model has a very promising clinical application.Based on the EchoNet-Dynamic public dataset,a study of left ventricular segmentation in echocardiography was performed.Several different semantic segmentation frameworks are used and the feature extraction network in the framework is replaced to achieve the segmentation of the left ventricle.And the Transformer-based semantic segmentation model is used for the segmentation of this data.In addition,the single-plane area method was used to estimate the ejection fraction,and the patients were graded according to the estimated ejection fraction.In this paper,two grading methods were used to evaluate the clinical performance of the models.To evaluate the comprehensive performance of each model,scoring rules were set up to evaluate the technical performance and clinical performance of the models separately in this paper.The results show that the B3 version of SegFormer achieves the best technical and clinical performance,which also demonstrates the great potential of the Transformer model applied to echocardiography to achieve real-time segmentation.