Uncertainty Quantification for Cardiac Diffusion Tensor Imaging Without Additional Datasets

Cardiac diffusion tensor imaging (cDTI) is subject to physiological noise, thermal noise, and signal corruption, which cause errors in diffusion measures. While a larger dataset can be decimated to investigate the general precision of measures from fitting smaller datasets, uncertainty quantification (UQ) methods for fitting entire particular datasets are required for UQ to be output from cDTI post-processing pipelines.

To account for non-idealized errors in cDTI, repetition bootstrap methods with whole-image resampling are required to approximate the sampling distribution of measures. We demonstrate UQ of voxel-wise diffusion measures and myocardial summary statistics over multiple voxels, as well as uncertainty-weighted summary statistics and their uncertainties. Methods are demonstrated on datasets of healthy volunteers and hypertrophic cardiomyopathy patients.

Group differences are larger (and p values smaller) for MD, FA and $\mid E2A\mid$ when myocardial averages of diffusion measures are weighted by uncertainty. This is particularly true for $\mid E2A\mid$ (difference of group medians: ${24.0}^{°}$ for unweighted average, ${36.7}^{°}$ for uncertainty weighted average). The uncertainty of averages over myocardial voxels is useful to understand outlier cases where it is difficult to determine if the result is trustworthy from diffusion measures alone. Uncertainty maps are also useful for highlighting regions of less trustworthy diffusion measures.

Uncertainty quantification in cardiac diffusion tensor imaging can be performed with respect to the sampling distribution of the available cDTI dataset, provided the dataset design is suitable for repetition bootstrapping.