The Segmentation Of Cardiac Region And Analysis Of Fibrosis Based On Deep Learning

T1 mapping technology is a noninvasive imaging method of Cardiac Magnetic Resonance(CMR).It can be used to directly evaluate the T1 value of myocardial tissue.We can find a variety of cardiovascular disease by the T1 mapping images such as myocardial infarction,non-ischemic cardiomyopathy,myocardial amyloidosis and so on.With the rapid development of CMR and T1 mapping technology,they are becoming more and more common in clinical applications.But current T1 mapping technology lacks uniform technical standard.The T1 values and the myocardial Extra Cellular Volume fraction would be different because of the different magnetic field intensity,different scanning sequence,different post-processing methods and some individual factors of patients themselves.So the interpretation of T1 mapping images depends on doctors and radiological technologists.They find the location of myocardium on T1 mapping images and judge whether there is abnormality.It is a heavy work and needs a lot of training for doctors.The missed diagnosis will directly delay the treatment opportunity of patients.It is not good for the patients.According to the problems existing in CMR image processing,we designed a method based on deep learning to segment and statistics myocardial tissue.The goal is to recognize the position and structure of the heart in the T1 mapping images and to analyze the degree of cardiac fibrosis.The main works of this paper are as follows.(1)We studied the application of the Yolov3 network to recognize the region of interest(ROI)of the heart.We used all MR images,which are used to calculate the T1 mapping images,to improve the whole detection rate.The experiment results showed that the detection rate of cardiac ROI recognition in groups reached 100%.(2)We segmented the myocardium in the ROI of CMR images by the Unet network with the Res Net-50 as the backbone.We propose a multi-image segmentation fusion method to improve the accuracy of myocardial segmentation in T1 mapping images.It fused myocardial segmentation masks of multiple CMR images.Comparing with other image segmentation models,the Res Net50＿Unet model with multi-image segmentation fusion achieves the best mean Dice of0.813 on the validation set.In order to obtain the fusion strategy needed by the multi-image segmentation fusion method,we study a fusion strategy acquisition method based on the Segfuse model,and its effect close to the artificial fusion strategy on the validation set.(3)The 16-segment division method from the American Heart Association was implemented and improved to divide the myocardium into 36 fine-grained regions.It makes easy to statistics of T1 values in T1 mapping images so that we can analyze the degree of fibrosis.The direction of the heart has a great influence on the determination of the upper and lower ventricular joint points.We determine the starting line of the division by the upper joint point and the centre point of the heart.So we enhance the training set for Yolov3 by rotating.It not only improves the performance of the Yolov3 network,but also can make Yolov3 distinguish the different directions of heart.This improvement greatly simplifies the myocardial fine-grained division algorithm.