Research On Cardiac MRI Segmentation Algorithm Based On Semi-Supervised Learning

Atrial fibrillation is a common persistent arrhythmia disease.As patients age,it easily leads to blood pooling in the atria and generates thrombus,which causing fatal complications such as heart failure and cerebral embolism.These kinds of complications can seriously threaten patients’ lives.Especially,clinical diagnosis using left atrial structure segmented by late gadolinium enhancement magnetic resonance imaging(LGEMRI)is the mainstream treatment.Although reconstructing the atrial structure by manual segmentation can effectively assist physicians in treatment,which requires a large amount of labor cost.Therefore,using deep learning methods for segmentation to reduce the reliance on experts in the field is an important research direction.In this paper,a semisupervised deep learning approach based on the above research background is used to investigate the fully automated segmentation algorithm of the left atrium for LGE-MRI.The main research contents and results are as follows.(1)A semi-supervised geometrically constrained consistent regularized left atrial segmentation method is proposed.The geometrically constrained consistency regularization method is designed to improve the accuracy of pixel-wise segmentation by converting the pixel-wise segmentation task to the geometric-wise through the level set function.And then,using the difference between the pixel-wise segmentation probability map and the level set segmentation map as a perturbation to design a consistency loss function.Under experimental conditions using only 20% labeled data and 80% unlabeled data for training,the highest segmentation accuracy of 89.47% is achieved,which is 3.4%better than the basic VNet method.It demonstrates that the geometrically constrained semi-supervised segmentation method can achieve accurate 3D left atrial segmentation using unlabeled data.(2)The use of mutual learning mechanism and optimized VNet network structure are proposed to enhance the training effect of geometrically constrained consistent regularization.Mutual learning knowledge distillation method differs from the traditional teacher-student model.Instead of just unilateral knowledge transfer from the teacher network to the student,two partner networks of the same level are trained together.And knowledge is transferred in both directions during the training process using the consistency loss function.This idea is exactly in line with the need for the geometrically constrained consistency regularization method to perform both pixel-wise segmentation prediction and level-set segmentation prediction.It means to have the two partner networks trained for pixel-level segmentation and level-set segmentation,respectively,and use pixel-level segmentation as the prediction result.As to optimized VNet,while retaining the VNet infrastructure,we add the Selective Kernel in the decoder;the atrous spatial pyramid pooling(ASPP)in the bottleneck layer;the self-attention mechanism in the jump connection;and the deep supervision mechanism in the upsampling layer of the decoder to improve the model’s ability to extract contextual features and restore deep features.In the same experimental setting,the geometrically constrained segmentation method incorporating mutual learning and optimized VNet can achieve 90.13%segmentation accuracy for LGE-MRI images of the left atrium,which is an additional0.66% improvement from the original accuracy.Finally,the contribution of each optimization strategy to the segmentation accuracy improvement is demonstrated by ablation experiments.