Unsupervised Quality Control of Image Segmentation Based on Bayesian Learning

Assessing the quality of segmentations on an image database is required as many downstream clinical applications are based on segmentation results. For large databases, this quality assessment becomes tedious for a human expert and therefore some automation of this task is necessary. In this paper, we introduce a novel unsupervised approach to assist the quality control of image segmentations by measuring their adequacy with segmentations produced by a generic probabilistic model. To this end, we introduce a new segmentation model combining intensity and a spatial prior defined through a combination of spatially smooth kernels. The tractability of the approach is obtained by solving a type-II maximum likelihood which directly estimates hyperparameters. Assessing the quality of the segmentation with respect to the probabilistic model allows to detect the most challenging cases inside a dataset. This approach was evaluated on the BRATS 2017 and ACDC datasets showing its relevance for quality control assessment.