DEEP LEARNING-BASED LEFT HEART CHAMBERS SEGMENTATION AND STRAIN ANALYSIS FROM DYNAMIC MRI IMAGES

Feature tracking (FT) is increasingly used on dynamic magnetic resonance images for myocardial strain evaluation, but often requires manual initialization of heart chambers, which is tedious and source of variability, especially on the challenging long axis images. Accordingly, we combined a deep learning (DL) approach with FT (DL-FT) to provide fully automated time-resolved left ventricular (LV) and atrial (LA) delineation and strain analysis. This approach was tested on a multi-center and multi-vendor database of 684 healthy controls and patients. DL-initialization achieved Dice scores of 0.89 +/- 0.11 for LV endocardium, 0.93 +/- 0.07 for LV epicardium and 0.89 +/- 0.10 for LA on the testing set of 108 datasets (2-and 4-chambers). LA and LV DL-FT strain peaks were highly associated with expert strains as revealed by correlation coefficients=0.96 for LV and >= 0.70 for LA and mean Bland-Altman biases=0.62% for LV and <1% for LA. Results also revealed stability of our approach over vendors and field strengths.