Shape Analysis of Segmentation Variability

Jess Tate1, Nejib Zemzemi2, Wilson Good3, Peter van Dam4, Dana Brooks5, Rob MacLeod1
1University of Utah, 2Inria Bordeaux Sud-Ouest, 3The SCI Institute, 4UMC Utrecht, 5Northeastern University


Patient-specific cardiac simulation is becoming increasingly relevant as a research and clinical tool for predicting arrhythmias and guiding treatments. Many simulation pipelines rely on accurate geometric models extracted from medical images. Segmentation of car- diac images is a key, yet possibly error-prone part of patient-specific simulations, e.g., heart propagation models, ECG forward simulation, and ECG Imaging. In this study, we performed shape analysis on multiple segmentations of the same patient to quantify vari- ability. We used the ShapeWorks software tool to find a mean shape, modes of shape vari- ation, and the extremes of each mode from multiple ventricular segmentations. Through a collaborative effort within the Consortium for ECG Imaging (CEI), a single patient CT scan was segmented by five research groups. Each segmentation included left and right ventricular endocardium and the pericardium. We used ShapeWorks to compute corre- spondence points on all surfaces based on an entropy optimization scheme. Principle component analysis (PCA) of the point locations determined the characteristics of the shape space. We found that segmentation shape varied most in the basal region of the ventricles in all three structures. Other areas of significant variation include near the RV apex and the inferior pericardium. The choice of how much of the valves and the outflow tracks to include in the segmentation contributed most to shape variability. The primary mode of variation contained variability of up to 43 mm in possible surface locations. Our results demonstrate a substantial variability in the segmentation of the ventricles, which could have significant impact on pipelines that depend on geometric models. The quantification of the shape variability using ShapeWorks provides a pathway to subse- quently quantify the impact of the segmentation variability on modeling pipelines with uncertainty quantification.