Cross-center smoothness prior for Bayesian image segmentation

Abstract

MR images vary across medical centers due to calibration and acquisition protocols. But segmentations are relatively consistent. How segmentations are supposed to look like, can be learned separately. We present a smoothness prior that is fit to segmentations from a source medical center. This empirical prior is incorporated into an unsupervised Bayesian image segmentation model. The model clusters voxel intensities in the target center, such that its segmentations are similarly smooth.