Research On Cardiac Magnetic Resonance Image Segmentation Algorithm Based On Capsule Network

In recent years,the mortality rate of cardiovascular diseases has been increasing year by year,which has seriously affected the life quality and health of the people.Therefore,it is of great significance to study the early auxiliary diagnosis of cardiovascular diseases.Magnetic Resonance Imaging(MRI)is an important means of inspection for cardiovascular disease.Cardiac magnetic resonance image segmentation is the basis of cardiac function index calculation and can provide reference information for clinical auxiliary diagnosis of cardiovascular diseases.the indexes of cardiac function can be calculated,which can provide quantitative reference information for clinical diagnosis.Due to fuzzy boundary,uneven gray distribution and other factors,the existing MRI image segmentation algorithm is not highly automated and robust to noise is poor.On the basis of analyzing the shortcomings of existing segmentation algorithms,a cardiac magnetic resonance image segmentation model based on Capsule network is proposed in this thesis.The main research work of this paper is as follows:(1)Aiming at the problem of incomplete image labeling in MICCAI2009 dataset,this thesis proposes a Level set based semi-automatic cardiac magnetic resonance image labeling model to realize the tagging of cardiac magnetic resonance image sequence and to provide a complete and high-quality data set for the training model.In this model,0 level set is used to represent the myocardial intima,k level set is used to represent the myocardial outer membrane,and the distance between the two level sets is constrained.Based on this,the energy function of ventricular label was established to realize the labeling of cardiac magnetic resonance images by minimizing the energy function.The Experiments show that the model proposed in this thesis completes the annotation of cardiac magnetic resonance images and solves the problem of incomplete image annotation in MICCAI2009 dataset.(2)Aiming at the problem that deep neural network can not model the inner and outer myocardium simultaneously,this thesis proposes a cardiac magnetic resonance image segmentation model based on Capsule network.The model based on the Capsule network,uses the parallel independence between the capsules of the Capsule network,extracts image information through nerve capsules,and adds the whole fully connected layer and de-convolution layer to reach the segmentation goal.The experimental results indicate that the model proposed in this thesis achieves efficient segmentation of the inner and outer myocardium and the Dice coefficient can reach 0.9417.When we add different gaussian noises,the Dice coefficient remains about 0.93,showing that the model proposed in this paper has great noise robustness.This thesis proposes a Level set-based semi-automatic cardiac magnetic resonance image labeling algorithm that provides a complete and high-quality data sets for the training model.On the basis of gaining complete data sets,this paper proposes a cardiac magnetic resonance image segmentation model based on Capsule network.It is pretty important for the auxiliary diagnosis of cardiovascular diseases.