Research And Application Of Ventricular Segmentation Using Nuclear Magnetic Resonance Imaging Based On Deep Learning

In recent years,Cardiovascular Disease(CVD)has caused serious threats to human health and has aroused widespread concern.At present,the diagnosis of cardiovascular disease is mainly completed by multi-slice spiral CT,echocardiography and nuclear magnetic resonance.Nuclear magnetic resonance is widely used in medical applications due to its characteristics of high temporal resolution,none radiation and clear imaging.However,manual division of nuclear magnetic images are time-consuming and error-prone,so automated segmentation is necessary.Ventricular endocardium markers provide an important basis for ventricular volume calculation,ej ection fraction and other indicators,researchers continuously explore the ventricular endocardium boundary detection in practice to improve the accuracy of cardiac function index calculation.Based on the short-axis ventricular magnetic slice,the study of ventricular endocardium segmentation was carried out.The main research contents are as follows:1.In view of the fact that there are few samples of open magnetic datasets in ventricular magnetic images,so 16,000 ventricular short-axis nuclear magnetic images of 128 patients were collected from a hospital of Gansu Province.The left and right ventricular endocardium boundaries were labeled one by one to produce a large dataset.2.Nuclear magnetic images have the characteristics of uneven distribution of gradation and the low contrast of ventricular endocardium boundary,so this thesis uses bilateral filtering and adaptive histogram equalization algorithm to improve data quality,which provides a foundation for accurate marking and segmentation of images.3.This thesis studies the hierarchical design model based on ventricular nuclear magnetic section,and then makes a combination of hierarchical design model and the deep learning-based Mask R-CNN algorithm to detect ventricular endocardium.The test set includes all the images in a cardiac cycle.The experimental results show that the mean DM(Dice Metric)coefficients of the left and right ventricles are 0.92 and 0.89,and the HD(Hausdorff Distance)coefficients are 4.78 mm and 7.03 mm.4.On the basis of the above studies,the proportional relation between end-diastole and the end-systole areas of the nuclear magnetic section in the second layer of the ventricle(near the tricuspid valve)was studied,and the persons with normal ventricular systolic function and abnormal ventricular systolic function were separated by this feature.In summary,the approach proposed in this thesis can automatically segment the ventricular endocardium through the combination of hierarchical design and Mask R-CNN algorithm,which achieves good results and provides a certain reference for future research work.