Cardiac Motion Tracking In MRI Images Based On Unsupervised Deep Learning

The incidence of cardiovascular disease in China is continuously increasing year by year,and the mortality rate ranks the first among various diseases,some of which are closely related to abnormal cardiac motion function.On the basis of accurately tracking the trajectory of the cardiac motion,the cardiac motion function can be analyzed and the cardiac disease can be diagnosed.The traditional cardiac motion tracking method is slow,the effect is not outstanding,which is not suitable for clinical application.The classification of heart diseases also requires manual extraction of cardiac morphological features,which is not convenient for computeraided diagnosis.In recent years,the wide application of deep learning,especially convolutional neural networks in the field of computer vision,has brought new solution to heart motion tracking and heart disease classification problems.Convolutional neural networks extract effective features for specific tasks based on feature learning,and achieve excellent results.In this thesis,the powerful feature extraction capabilities of convolutional neural networks are applied to the fields of cardiac motion tracking and cardiac disease classification.Aiming at the problem of cardiac motion tracking,this dissertation proposes a method for cardiac motion tracking based on unsupervised learning.This method proposes to significantly reduce the number of network parameters by reasonably reducing the number of network channels,and adds the Multi-resolution Feature Aggregation Module for improvement,resulting in the efficient Multi-resolution Feature Aggregation Network as subnet.Based on the idea of global and local heart tracking task allocation,a dual network architecture was designed,the main network is mainly responsible for the prediction of the overall motion mode of the heart,the secondary network is mainly responsible for correcting the local motion of the heart,using multi-level loss function constraints to ensure the task distribution of the primary and secondary networks,and finally synthesizing the motion trajectories obtained by the primary and secondary networks to get a complete cardiac motion trajectory.The proposed network structure is called the Residual Dual Network,the left ventricular,right ventricular and myocardial motion tracking dice coefficients are 0.948,0.903,0.856 in the MICCAI-ACDC dataset,and the model parameters are only 7M,which achieves a balance between performance and speed.For the classification of cardiac diseases,this thesis using the results of cardiac motion tracking to propose the Myocardial Infarction Diagnosis Two-stream Network based on convolutional neural networks.The network contains two sub-network branches for image information extraction and motion information extraction,which are used to process static information such as morphology and position contained in the image and dynamic information contained in the motion trajectory.Each branch contains a feature compression module for mapping 3D data into 2D features and a feature encoder module for classifying features,the final classification result is the fusion of two sub-branch outputs.This thesis presents an endto-end cardiac disease classification algorithm,without manually extracting features,and the average classification accuracy of the network is 0.772.The Myocardial Infarction Slice Diagnosis Two-stream Network is further proposed to classify normal cardiac slices and slices where exist myocardial infarction areas,which can locate the slice of myocardial infarction in patient’s heart and judge the severity of myocardial infarction.The average classification accuracy of this method for normal slices and myocardial infarction slices is 0.816.