Automatic segmentation of left ventricle from cardiac MRI via deep learning and region constrained dynamic programming

Abstract

Segmentation of the left ventricle from cardiac magnetic resonance images (MRI) is an essential step to quantitatively analyze global and regional cardiac function. The aim of this study is to develop a novel and robust algorithm which can improve the accuracy of automatic left ventricle segmentation on short-axis cardiac MRI. The database used in this study are 900 cardiac MRI cases from Hubei Cancer Hospital. Three key techniques are developed in this segmentation algorithm: (1) deep learning schema is designed for coarse segmentation of LV (left ventricle) myocardial images, (2) trimmed mean approach is employed to build a proper region for epicardial contour searching, and (3) an edge map with non-maxima gradient suppression approach is put forward to improve the dynamic programming to derive the epicardial boundary. The accuracy of cardiac boundaries and cardiac functional parameters are benchmarked against those derived from expert manual boundary delineations and derived measurements on 900 samples of Hubei Cancer Hospital. The average Jaccard (overlap) indexes for endo and epicardial contours are 0.80 and 0.76. Bland-Altman analysis for LVEDV (LV End-diastolic Volume), LVESV (LV End-systolic Volume), LVSV (LV 232 Stroke volume), LVEF (LV Ejection Fraction), LVM-DP (LV mass in diastolic phase) and LVM-SP (LV mass in systolic phase), yielded biases of 4.15 ml, 0.26 ml, 3.89 ml, 0.32%, −7.24 g and -0.08 g, respectively, and limits of agreement of ±30.32 ml, ±11.10 ml, ±24.31 ml, ±12.89%, ±24.66 g and ±21.93 g, respectively. Our hybrid scheme combining deep learning and dynamic programming for automatic segmentation of the LV from cardiac MRI datasets is effective and robust and can compute LV functional indexes from population imaging.