Automatic delineation of left ventricle contours in MRI using deep learning

Delineation of Left Ventricle (LV) contours is a common task in the clinical diagnosis of cardiac abnormalities. This step is crucial, especially in Cardiac Magnetic Resonance Imaging (CMRI), as it allows to estimate important parameters for the quantification of cardiac contractile function such as cardiac volumes allowing calculation of the Ejection Fraction (EF) and myocardial thickening. The aim of this paper is to optimize and adapt the standard U-net architecture to our 2D cardiac Cine-MRI database in order to perform LV contours extraction automatically. The proposed method has been validated and tested on 70 patients with various cardiac pathologies. The obtained results show a great concordance between the ground truth and predicted segmentation with a mean Dice Similarity Coefficient (DSC) and Hausdorff Distance (HD) which reach 97.22% and 6.192 mm respectively at the end diastolic phase. These metric coefficients reach 93.5% and 8.679 mm respectively at the end systolic phase. The comparison results prove that the adopted method is likely to be considered as an effective tool for fully automatic LV segmentation.