A scalable method to improve gray matter segmentation at ultra high field MRI

Abstract

High-resolution (functional) magnetic resonance imaging (MRI) at ultra high magnetic fields (7 Tesla and above) enables researchers to study how anatomical and functional properties change within the cortical ribbon, along surfaces and across cortical depths. These studies require an accurate delineation of the gray matter ribbon, which often suffers from inclusion of blood vessels, dura mater and other non-brain tissue. Residual segmentation errors are commonly corrected by browsing the data slice-by-slice and manually changing labels. This task becomes increasingly laborious and prone to error at higher resolutions since both work and error scale with the number of voxels. Here we show that mislabeled, non-brain voxels can be corrected more efficiently and semi-automatically by representing three-dimensional anatomical images using two-dimensional histograms. We propose both a uni-modal (based on first spatial derivative) and multi-modal (based on compositional data analysis) approach to this representation and quantify the benefits in 7 Tesla MRI data of nine volunteers. We present an openly accessible Python implementation of these approaches and demonstrate that editing cortical segmentations using two-dimensional histogram representations aids existing algorithms and yields improved gray matter borders. By making our data and corresponding expert (ground truth) segmentations available, we facilitate future efforts to develop and test segmentation algorithms on this challenging type of data.

CEREBRUM-7T: Fast and Fully-volumetric Brain Segmentation of 7 Tesla MR Volumes

“…As a consequence, manual segmentation protocols (Wenger et al, 2014;Berron et al, 2017), although time consuming (Zhan et al, 2018;Koizumi et al, 2019), are still a common practise for 7T data. To partially reduce the laboursome process of manual segmentation, the solution proposed by Gulban et al (2018) combines manual and semi-automatic segmentation, by adopting a multi-dimensional transfer function to single out non-brain tissue voxels in 7T MRI data of nine volunteers. Other semi-automated methods developed in the past for generic MRI data, such as ITK-SNAP (Yushkevich et al, 2006), have been adapted by (Archila-Meléndez et al, 2018) for tackling also ultra-high field brain imaging.…”

“…The method, which includes a non-trivial preprocessing chain for skull stripping and dura estimation, achieves whole brain segmentation and cortical extraction, all within a computation time below six hours. Despite these efforts, most existing solutions, including Bazin et al (2014), still generate a variety of segmentation errors that needs to be manually addressed, as reported in Gulban et al (2018).…”

LayNii: A software suite for layer-fMRI

“…E.g. the interested reader is referred to an instruction of how to use FreeSurfer, nighres and SUMA to generate an input rim file for LayNii here: (https://layerfmri.com/getting-layers-in-epi-space/), this segmentation can be further corrected with the semi-manual segmentation tool Segmentator (Gulban et al 2018) (https: //github.com/ofgulban/segmentator).…”

Imaging the human auditory system at ultrahigh magnetic fields

“…We have implemented the methods described here in a free and open Python software package . The package as well as validation data, corresponding expert segmentations (Gulban, Schneider, Marquardt, Haast, & De Martino, 2017), and processing scripts used to validate the proposed methods are all openly available (see Table 2.1 for links).…”