Assessment of Cardiac Allograft Vasculopathy from Oct Images: Automated Analysis is as Good as Expert-Guided Approach.
Researchers
Journal
Modalities
Models
Abstract
Cardiac Allograft Vasculopathy (CAV) is a frequent complication after heart transplantation (HTx). To help identify patients at risk of CAV, 3D quantitative analysis of coronary wall thickening is of major importance. Until now, substantial manual tracing effort was required. We report a fully automated approach using optical coherence tomography (OCT) imaging.
Lumen surface, intimal and medial layers were identified using our LOGISMOS segmentation framework. Coronary wall regions with layered appearance were automatically identified in each OCT frame using deep learning. These segmentation and classification approaches were newly combined in one fully automated system. The comparison between fully automated and expert-guided analyses was performed in 116 patients with OCT imaging at 1 and 12 months after HTx (1M, 12M). The expert-guided analysis used our previously-reported semi-automated layer segmentation method with expert-identified analyzable regions thus yielding an independent standard.
The table shows that our fully automated approach yields quantitative indices of CAV progression that are statistically indistinguishable from the independent standard using Wilcoxon signed-rank test. The observed measurement errors are very small fractions of the OCT pixel size (∼10 μm/pixel). While the expert-guided definition of the independent standard required 2-15 minutes of user-guided segmentation and at least 30 minutes of manual definition of layered wall regions per pullback, the new automated analysis required less than 2 minutes of computer processing for each coronary pullback (median length 38.7 mm).
Our approach is fully automated, efficient, and facilitates accurate layer-specific quantification of coronary wall thickness and their changes 1M-12M after HTx. The fully automated method can replace the manual or expert-guided analyses of OCT image sequences while maintaining high levels of accuracy.
Copyright © 2020. Published by Elsevier Inc.