Corrected Uncertainty in Probabilistic Segmentation Using Local Statistics

Abstract

Probabilistic segmentation algorithms compute for each voxel and each segment of a medical imaging data set a probability that the voxel belongs to the segment. These per-voxel probability vectors are commonly used to estimate uncertainties and produce respective visualizations. It can be observed that one obtains high uncertainties along the border of two adjacent tissues, even in case of high gradients. This is due to the partial volume effect (PVE). PVE, however, is not a source of uncertainty. In case of high-gradient borders, one can be very certain that respective voxels partially belong to one and partially to the other voxel. We correct this misconception by modeling PVE using local statistics within a probabilistic segmentation approach. As a result we obtain corrected uncertainties and we even have been able to significantly improve the probabilistic segmentation approach itself.