A key aspect of cardiac modeling is the creation of a patient-specific torso model by segmenting radiological images, a subjective and error-prone process. Our previous work has sought to quantify the effect of segmentation variability on cardiac models, but those studies did not explore activation sequences that originate from regions where segmentation variability is particularly high. In this study, we quantified the uncertainty in cardiac forward models resulting from cardiac segmentation variability, focusing on how the uncertainty varies with activation sequence.
We applied our shape uncertainty modeling pipeline, based on statistical shape modeling and polynomial Chaos (PC), to cardiac forward models. The shape model was constructed from 15 segmentations of the same subject, then sampled and evaluated in two forward modeling pipelines: an Eikonal propagation model and a nearest neighbor propagation model. We used PC to estimate the statistics of the resulting model variation.
The quantified model uncertainty due to cardiac shape variability was higher in regions of high segmentation variability. Activation sequences originating from high shape variability regions, notably the base of the heart and the right ventricular outflow track (RVOT), showed higher uncertainty than other activation sequences which was high enough to generate clinically significant QRS morphology changes. Pacing near the base showed higher uncertainty than near the apex, and RV pacing showed higher uncertainty than LV.
Our results provide further insight into the response of ECG forward models to segmentation variability. While cardiac forward models can be relatively robust to shape variability in many circumstances, in some applications, such as modeling pacing sites originating where segmentations can be highly variable, model solutions can be highly sensitive to the variability. These errors can become compounded when used to feed further computations, such as when modeling data is used to develop ECGI pipelines.