Direct cortical thickness estimation using deep learning-based anatomy segmentation and cortex parcellation

Abstract

Accurate and reliable measures of cortical thickness from magnetic resonance imaging are an important biomarker to study neurodegenerative and neurological disorders. Diffeomorphic registration-based cortical thickness (DiReCT) is a known technique to derive such measures from non-surface-based volumetric tissue maps. ANTs provides an open-source method for estimating cortical thickness, derived by applying DiReCT to an atlas-based segmentation. In this paper, we propose DL+DiReCT, a method using high-quality deep learning-based neuroanatomy segmentations followed by DiReCT, yielding accurate and reliable cortical thickness measures in a short time. We evaluate the methods on two independent datasets and compare the results against surface-based measures from FreeSurfer. Good correlation of DL+DiReCT with FreeSurfer was observed (r = .887) for global mean cortical thickness compared to ANTs versus FreeSurfer (r = .608). Experiments suggest that both DiReCT-based methods had higher sensitivity to changes in cortical thickness than Freesurfer. However, while ANTs showed low scan-rescan robustness, DL+DiReCT showed similar robustness to Freesurfer. Effect-sizes for group-wise differences of healthy controls compared to individuals with dementia were highest with the deep learning-based segmentation. DL+DiReCT is a promising combination of a deep learning-based method with a traditional registration technique to detect subtle changes in cortical thickness.

Keywords: MRI; brain morphometry; cortical thickness; deep learning; diffeomorphic registration; gray matter atrophy; neuroanatomy segmentation.

FIGURE 1

DL+DiReCT: Deep learning‐based neuroanatomy segmentation followed by a diffeomorphic registration to estimate cortical thickness from MRI

FIGURE 2

Three samples (one per row) from the OASIS‐3 dataset. Columns show T1‐weighted MRI with pial (blue) and GM/WM (yellow) surface from FreeSurfer overlayed, segmentations from FreeSurfer and deep learning (DL), and thickness map from DL+DiReCT. Slices are in radiological view (i.e., right hemisphere is on the left side of the image)

FIGURE 3

Color‐coded Pearson correlation coefficients (r) of the ROI‐wise average cortical thicknesses compared to FreeSurfer evaluated on the OASIS‐3 samples

FIGURE 4

Comparison of the global mean thickness estimations against FreeSurfer (FS) for DL+DiReCT (first row) and ANTs (second row) for the samples in the OASIS‐3 dataset. Left: correlation plot. Middle: Bland–Altman plot, dashed horizontal line indicating ±1.96σ. Right: Thicknesses plotted against age

FIGURE 5

Color‐coded reproducibility errors of the ROI‐wise average cortical thicknesses evaluated on the OASIS‐3 samples

FIGURE 6

Color‐coded annual atrophy rates in mm/year of the ROI‐wise average cortical thicknesses evaluated on the OASIS‐3 samples

FIGURE 7

Kernel density plots of the global mean thickness, corrected for brain size and age, depicting effect‐size (Cohen's d reported in the subtitle) between healthy controls (HC) and dementia