Study Of Fully Automatic Myocardial Segmentation Algorithm Based On Deep Learning Methods

Myocardial segmentation plays an important role in the clinical diagnosis of cardiomyopathies.By using the results of myocardial segmentation,it is easy to perform three-dimensional modeling on the heart and calculate the myocardial thickness and myocardial mass.Myocardial segmentation is also the basis of many cardiac image processing,such as registration,map construction and so on.Due to the characteristics of cardiac muscle tissue such as unclear boundary,large difference in myocardial structure between different slice locations and different time series,myocardial segmentation has always been a challenging task in heart image processing,and the traditional image segmentation algorithm is difficult to have good segmentation accuracy and robustness in myocardial segmentation tasks.The traditional image segmentation algorithm is usually semi-automatic segmentation algorithm,requiring human involvement,which brings additional workload.Therefore,it is necessary to propose a fully automatic,highly accurate and robust myocardial segmentation algorithm.After analyzing the deficiencies of existing myocardial segmentation algorithms both at home and abroad,in this paper we propose two kinds of fully automatic myocardial segmentation algorithms based on deep learning methods with different frameworks:Myocardial segmentation algorithm based on image block classification and Myocardial segmentation algorithm based on fully convolutional networks.Among them,the myocardial segmentation algorithm based on image block classification improves the traditional algorithm of segmentation based on image block classification,which is time-consuming and has large amount of computation,by the cooperation of cardiac localization network and myocardial segmentation network.On the test set,this method achieves an average of 90.23%of the DSC results with an average speed of 0.94s per image processing.And the myocardial segmentation algorithm based on fully convolutional networks greatly improves the myocardial segmentation efficiency by constructing an end-to-end neural network model,in which the HeartNet model we proposed achieves an average of 90.48%of the DSC results on the test set with an average speed of 144.9FPS.We evaluated our proposed method on Kaggle's Data Science Bowl Cardiac Challenge Data,which consists of 1140 patients' cine CMR images.For each patient,it contains approximately 30 images across the cardiac cycle in each short axis stack.By analyzing the experimental results and comparing with other segmentation algorithms,the accuracy and robustness of the proposed algorithm are verified,and the value of the algorithm in clinical application is confirmed.In this paper,we also built a fully automatic myocardial segmentation platform based on our proposed algorithm to improve the practicality of the algorithm.