New tissue priors for improved automated classification of subcortical brain structures on MRI

Abstract

Despite the constant improvement of algorithms for automated brain tissue classification, the accurate delineation of subcortical structures using magnetic resonance images (MRI) data remains challenging. The main difficulties arise from the low gray-white matter contrast of iron rich areas in T1-weighted (T1w) MRI data and from the lack of adequate priors for basal ganglia and thalamus. The most recent attempts to obtain such priors were based on cohorts with limited size that included subjects in a narrow age range, failing to account for age-related gray-white matter contrast changes. Aiming to improve the anatomical plausibility of automated brain tissue classification from T1w data, we have created new tissue probability maps for subcortical gray matter regions. Supported by atlas-derived spatial information, raters manually labeled subcortical structures in a cohort of healthy subjects using magnetization transfer saturation and R2* MRI maps, which feature optimal gray-white matter contrast in these areas. After assessment of inter-rater variability, the new tissue priors were tested on T1w data within the framework of voxel-based morphometry. The automated detection of gray matter in subcortical areas with our new probability maps was more anatomically plausible compared to the one derived with currently available priors. We provide evidence that the improved delineation compensates age-related bias in the segmentation of iron rich subcortical regions. The new tissue priors, allowing robust detection of basal ganglia and thalamus, have the potential to enhance the sensitivity of voxel-based morphometry in both healthy and diseased brains.

Keywords: Basal ganglia; Effective transverse relaxation; Magnetization transfer saturation; Relaxometry; Tissue probability maps; Voxel-based morphometry; Voxel-based quantification.

Fig. 1

Flowchart summarizing the steps to create the new tissue probability maps (TPM). (a) Atlas information and manual labeling of binary masks relative to subcortical structures. (b) Creation of the new TPM from the binary masks.

Fig. 2

Top panel: axial view on the new and the conventional tissue probability maps (TPMs) for gray matter. Bottom panel: example of tissue classification from T1-weighted data using the new and the conventional TPMs. The gray matter probability maps is represented in red, the white matter—in yellow and the cerebro-spinal fluid (CSF)—in blue.

Fig. 3

Statistical parametric maps of paired t-test between gray matter (GM) volumes estimated from T1-weighted data with new and conventional tissue probability maps (TPM) at statistical threshold of p_FWE < 0.05. (a) Increase of GM volume estimation, based on the new TPM, compared to the estimation based on conventional TPM. (b) Increase of GM volume estimation, based on the old TPM, compared to the estimation based on conventional TPM. The pie charts represent the effect size for the indicated brain location. All results are presented at p_FWE < 0.05.

Fig. 4

Interaction between age-related gray matter (GM) volume loss and tissue probability maps (TPM) used for GM volume estimation from T1-weighted data. Top panel: statistical parametric maps of stronger negative correlation between age and GM volume estimates from new TPM compared to old TPM. Bar plot—mean linear regressors for age effects on volume in the pallidum. Bottom panel: statistical parametric maps of stronger negative correlation between age and GM volume estimates from old TPM compared to the new TPM. Bar plot—mean linear regressors for age effects on volume in the putamen.

Fig. 5

Top panel: t-values of voxel-based regressors correlating R2* with higher age-related gray matter (GM) volume loss estimated with conventional tissue probability maps (TPM). Bottom panel: t-values for the positive linear correlation between R2* and age in GM voxels.