Towards Left Ventricle Segmentation From Magnetic Resonance Images

Quantitative evaluation of cardiac function from magnetic resonance images generally requires the clinician to first trace the left ventricle contours. However, detection of myocardial walls continues to remain a challenge on magnetic resonance images acquired from patients having serious pathologies because of three key issues: 1) low contrast; 2) high noise level; and 3) blood pool region in the left ventricle is highly non-homogeneous. In this paper, a semi-automatic graph-based method is proposed to segment such pathological left ventricles. This paper has mainly three contributions, and they are: 1) weighting function in graph-based approaches for image segmentation is thoroughly analyzed; 2) a new weighting function is introduced for graph-based methods to outline the endocardium; and 3) epicardium is extracted by a proposed active contour model. We have tested the algorithm on real data sets obtained from two sources, Hospital for Sick Children (SICK-KID), Toronto, and MICCAI Left Ventricle Segmentation Challenge. Average Dice coefficients (in %), false positive ratio, false negative ratio, sensitivity, and specificity for SICK-KID database are found to be 94.7 ± 1.1, 0.023, 8.13 × 10
^-3
, 0.93, and 0.76, respectively; for MICCAI database, they are 96.1 ± 1.1,0.022, 8.10 × 10
^-3
, 0.94, and 0.74, respectively. Average Hausdorff distances between segmented contour and ground truth in these two databases are determined to be 2.86 and 2.84 mm, respectively. Promising experimental results by the method tested on publicly available database demonstrate the potential of the approach.