Medical image segmentation is a challenging task and algorithms often struggle with the high variability of inhomogeneous clinical data, demanding different parameter settings or resulting in weak segmentation accuracy across different inputs. Assessing the uncertainty in the resulting segmentation therefore becomes crucial for both communicating with the end-user and calculating further metrics of interest based on it, for example, in tumor volumetry. In this paper, we quantify segmentation uncertainties in a energy minimisation method where computing probabilistic segmentations is non-trivial. We follow recently proposed work on random perturbation models that enables us to sample segmentations efficiently by repeatedly perturbing the energy function of the CRF followed by MAP inference. We conduct experiments on brain tumor segmentation, with both voxel and supervoxel perturbations, and demonstrate the benefits of probabilistic segmentations by means of precision-recall curves and uncertainties in tumor volumetry along time.     «

Medical image segmentation is a challenging task and algorithms often struggle with the high variability of inhomogeneous clinical data, demanding different parameter settings or resulting in weak segmentation accuracy across different inputs. Assessing the uncertainty in the resulting segmentation therefore becomes crucial for both communicating with the end-user and calculating further metrics of interest based on it, for example, in tumor volumetry. In this paper, we quantify segmentation...     »