MRI whole heart segmentation using discrete nonlinear registration and fast non-local fusion

Abstract

We present a robust and accurate method for multi-atlas segmentation of whole heart MRI scans. After preprocessing, which includes resampling to isotropic voxel sizes and cropping or padding to same dimensions, all training scans are registered linearly and nonlinearly to an unseen set of test scans. We employ the efficient discrete registration framework called deeds that captures large shape variations across scans, performed best in a recent registration comparison on abdominal scans and requires less than 2 min of computation time per scan. Subsequently, we perform multi-atlas label fusion using a non-local means approach with a normalised SSD metric and a fast implementation using boxfilters. Subsequently, a multi-label random walk is performed on the obtained probability maps for an edge-preserving smoothing. Without performing any domain-specific parameter tuning, we obtained a Dice accuracy of 86.0% (averaged across 7 labels) and 87.0% for the whole heart on the MRI test dataset, which is the first rank of the MICCAI 2017 challenge. The segmentations are also visually very smooth using this fully automatic method.