Automated segmentation of cardiac adipose tissue in Dixon magnetic resonance images

Abstract

Objective: Increasing evidence suggests a strong link between excess cardiac adipose tissue (CAT) and the risk of a cardiovascular event. Multi-echo Dixon magnetic resonance imaging (MRI), providing fat-only and water-only images, is a useful tool for quantification but requires the segmentation of CAT from a large number of images. The intent of this study was to evaluate an automated technique for CAT segmentation from Dixon MRI by comparing the contours identified by the automated algorithm to those manually traced by an observer. 

Methods: An automated segmentation algorithm, based on optimal thresholds and custom morphological processing, was applied to the registered fat-only and water-only images to identify CAT in the volume scans. CAT contours in 446 images, from 10 MRI scans, were selected for validation analysis. Cross-sectional area (CSA) and volume were computed and compared using Bland-Altman analysis. In addition, Hausdorff distance and Dice Similarity Coefficient (DSC) were used for assessment.

Results: Linear regression analysis yielded correlation of R² = 0.381 for CSA and R² = 0.879 for volume. When compared to the observer, the computer algorithm under-estimated CSA by 27.5 ± 40.0% and volume by 26.4 ± 10.4%. The average bidirectional Hausdorff distance was 26.2 ± 16.0 mm while the average unidirectional Hausdorff distances were 24.5 ± 15.7 mm and 12.4 ± 11.7 mm. The average DSC was 0.561 ± 0.100. The time required for manual tracing was 15.84 ± 3.73 min and the time required for the computer algorithm was 2.81 ± 0.12 min.

Conclusions: This study provided a technique, faster and less tedious than manual tracing (p < 0.00001), for quantification of CAT in Dixon MRI data, demonstrating feasibility of this approach for cardiac risk stratification.