Uncertainty estimation and visualization in probabilistic segmentation

Probabilistic segmentation is concerned with handling imperfections of image segmentation algorithms. They assign to each voxel and each segment a probability that the voxel belongs to the segment. This is often the starting point for estimating and visualizing uncertainties in the segmentation result. We propose a novel, generally applicable uncertainty estimation approach that considers all probabilities to compute a single uncertainty value between 0 and 1 for each voxel. It is based on aspects of information theory and uses the Kullback-Leibler divergence (or the total variation divergence). We developed several forms of the proposed approach and analyze and compare their behaviors. We show the advantage over existing approaches, derive aggregated uncertainty measures that are useful for judging the accuracy of a probabilistic segmentation algorithm, and present visualization methods to highlight uncertainties in segmentation results.