Texture-based probability mapping for automatic scar assessment in late gadolinium-enhanced cardiovascular magnetic resonance images

Abstract

Purpose To evaluate a novel texture-based probability mapping (TPM) method for scar size estimation in LGE-CMRI.

Methods This retrospective proof-of-concept study included chronic myocardial scars from 52 patients. The TPM was compared with three signal intensity-based methods: manual segmentation, full-width-half-maximum (FWHM), and 5-standard deviation (5-SD). TPM is generated using machine learning techniques, expressing the probability of scarring in pixels. The probability is derived by comparing the texture of the 3 × 3 pixel matrix surrounding each pixel with reference dictionaries from patients with established myocardial scars. The Sørensen-Dice coefficient was used to find the optimal TPM range. A non-parametric test was used to test the correlation between infarct size and remodeling parameters. Bland-Altman plots were performed to assess agreement among the methods.

Results The study included 52 patients (76.9% male; median age 64.5 years (54, 72.5)). A TPM range of 0.328–1.0 was found to be the optimal probability interval to predict scar size compared to manual segmentation, median dice (25th and 75th percentiles)): 0.69(0.42–0.81). There was no significant difference in the scar size between TPM and 5-SD. However, both 5-SD and TPM yielded larger scar sizes compared with FWHM (p < 0.001 and p = 0.002). There were strong correlations between scar size measured by TPM, and left ventricular ejection fraction (LVEF, r = −0.76, p < 0.001), left ventricular end-diastolic volume index (r = 0.73, p < 0.001), and left ventricular end-systolic volume index (r = 0.75, p < 0.001).

Conclusion The TPM method is comparable with current SI-based methods, both for the scar size assessment and the relationship with left ventricular remodeling when applied on LGE-CMRI.