Comparison of semi-automatic myocardial scar segmentation methods on black-blood late gadolinium enhancement images

Myocardial scar quantification is key for patient prognosis. Black-blood late gadolinium enhancement (LGE) imaging improves scar visualization by nulling both blood and myocardium, solving the blood-scar interface contrast issue of bright-blood LGE. With no consensus on semi-automatic methods, scar segmentation is often manual, and therefore time-consuming and operator-dependent. We evaluate semi-automatic scar segmentation algorithms on co-registered bright- and black-blood LGE images.

Twenty patients with ischemic heart disease were included. Breath-held short-axis 2D whole-heart joint bright- and black-blood SPOT and reference PSIR images were collected on a 1.5 T system in mid-diastole. Left ventricle wall and scar were manually segmented by two experts on both images. The n-SD (n from 2 to 15), FWHM and Otsu thresholding scar segmentation methods were applied to black-blood SPOT images. Dice Similarity Coefficient (DSC), scar extent, volume and transmurality were measured for all segmentations.

Manual scar segmentation DSCs are higher on SPOT than PSIR images at both intra- (DSC: 87.9%[85.1%-88.8%]) and interobserver (87.3% [79.4%-89.4%]) levels. SPOT and PSIR yielded similar scar extent and volumes with moderate to good agreement (0.59 < ICC < 0.77). The 6-SD achieved the highest DSC (81.3%[73.7%-76.8%]), the most accurate scar extent (bias(SD): -0.8%(9.6%)), scar volume (bias(SD): 0.5 mL(7.1 mL)) and scar transmurality (bias(SD): -0.29%(10.70%)) for 7-, 8- and 3-SD, respectively.

Conventional semi-automatic scar segmentation algorithms developed for bright-blood imaging can be applied to black-blood imaging. Validation in larger populations is needed to determine and refine the most accurate approach.