Automatic Segmentation of Abdominal MRI Using Selective Sampling and Random Walker

MRI segmentation is a challenging task due to low anatomical contrast and large inter-patient variation. We propose a featuredriven automatic segmentation framework, combining voxel-wise classification with a Random-Walker (RW) based spatial regularization. Typically, such steps are treated independently, i.e. classification outcome is maximized without taking into account the regularization to follow. Herein we present a method for selective sampling of training patches, in view of the posterior spatial regularization. This aims to concentrate training samples near desired anatomical boundaries, around which the gain from a subsequent RW regularization will potentially be minimal. This trades off a lower classification accuracy for a higher joint segmentation performance. We compare our proposed sampling strategy to conventional uniform sampling on 20 full-body MR T1 scans from the VISCERAL dataset, both with RW and Markov Random Fields regularizations, showing Dice improvements of up to 12× with the proposed approach.