Cardiac Ventricular Segmentation Using Deep Learning Network With Semantic Flow And Self-attention Mechanism

Accurate segmentation of the ventricular region of cardiac magnetic resonance images plays an important role in the diagnosis and treatment of clinical heart disease.However,manual segmentation of ventricular regions is an extremely time-consuming task,and the accuracy of manual segmentation is easily affected by the quality of images and inter-observer experience.Therefore,in order to address the characteristics of the left and right ventricular regions of cardiac magnetic resonance images and the segmentation difficulties,this paper carries out the research on the deep learning-based segmentation method of the left and right ventricular regions.The main research work and innovation points are summarized as follows.1.A new LV segmentation method(SemConvGRU-Net)based on Semantic Streaming Convolutional Gated Recurrent Unit(SemConvGRU)and compound loss function is proposed to address the role of current deep learning-based LV region segmentation models that only consider singleframe 2D image slice spatial features and ignore inter-frame temporal features.First,a convolutional gated recurrent unit(Conv GRU)is used to extract temporal and spatial features between continuous heart MRI slices in the downsampling stage;then,SemConvGRU is used to fuse deep highlevel semantic information and shallow spatio-temporal feature information in the upsampling stage to reduce the loss of LV boundary information in the deeper layers of the network;finally,a composite loss function with dice loss function and image gradient loss The composite loss function integrated by the dice loss function and the image gradient loss function is introduced to solve the imbalance of occupancy between the LV region and the background.Extensive validation and comparison experiments of the proposed model on the dataset provided by Toronto Children’s Hospital show that SemConvGRU-Net significantly outperforms the current segmentation model and the segmentation performance is very similar to the manual segmentation results of clinical experts.2.To address the problem that the RV region is severely unevenly occupied with the background region and has indistinguishable similar signal characteristics with adjacent organs and tissues within the cavity and myocardium,exhibiting fuzzy boundaries,irregular cavities,and complex crescentic structural characteristics,which make the RV region difficult to segment accurately,a novel RV segmentation method using a multi-scale feature aggregation with a self-attentive mechanism(AttFeaAgg-Net).The multi-scale features of cardiac MRI are extracted using a self-attentive multi-scale feature expansion block(Sa Ms FEB)and the internal RV features in the feature space domain and channel domain are dynamically focused by the self-attentive mechanism;and a composite loss function integrated with exponential log loss and focus loss is introduced to solve the problem of degraded model segmentation performance caused by the imbalance between the occupancy ratio of RV region and background in cardiac MRI.Extensive validation and comparative experiments of AttFeaAgg-Ne on the MICCAI dataset show that AttFeaAgg-Net has better segmentation performance compared with existing methods,and its segmentation performance is very similar to the manual segmentation results of clinical experts.