要旨
Introduction: The thalamus and its constituent nuclei are implicated in several neurological and neuropsychiatric disorders, but the roles of specific nuclei are still being elucidated as thalamic nuclei segmentation methods have been suboptimal, until recently. Thalamus optimized multi-atlas segmentation (THOMAS) is a recently developed state-of-the-art method [1] that utilizes the improved contrast of white-matter nulled (WMn) MRI to generate accurate nuclei parcellations. It has been validated against manual segmentation guided by the Morel atlas. Recently, THOMAS has been adapted for standard 3T T1 MRI [2] by synthesizing WMn-MRI data from standard T1 MRI prior to segmentation. Normative modeling [3,4] is an emerging framework that has proven sensitive in detecting subtle heterogeneity often present in mental disorders, missed by conventional case-control analyses. Using the modified THOMAS method, we analyzed thousands of T1 MRI datasets from publicly available databases and created what we believe to be the first normative model of thalamic nuclear volumes across the lifespan. We then used this normative model to examine data from patients with dementia (early and late mild cognitive impairment, Alzheimer's disease), ADHD, bipolar disorder, affective and non-affective psychosis, and schizophrenia. Methods: Volumes of the whole thalamus and 12 nuclei were generated for each side using THOMAS in control subjects (n=2374, 1279 female 1095 male) and disease cohorts (n=694, 301 female 393 male). Site effects were harmonized using COMBAT-GAM and volumes were adjusted by total intracranial volume (FreeSurfer eTIV output). Multiple modeling approaches have been presented for normative modeling. To determine the optimal model for thalamic nuclear volumes, three models- Ordinary least squares regression (OLS), multiple fractional polynomial models (MFP), and generalized additive models of location shape and scale (GAMLSS)- were evaluated with 5-fold cross validation on a subset of 200 subjects. Using maximum absolute error and mean squared error as metrics, MFP models performed the best and were computationally efficient (in concordance with [4]) and was used for creating NMs on the whole dataset. For each nucleus and hemisphere, an MFP-based NM was trained on control subjects in sex-stratified datasets with age as a covariate, resulting in 13*2*2=52 models. These models were then applied to patient data from the disease cohorts. For each control subject and patient, z-scores were calculated from model residuals. Infranormal (z<=-2) and supranormal (z>=2) deviations were tabulated. Multi group comparisons were performed using the Kruskal-Wallis rank sum test and z-score distributions were compared using the z-test. Results: Figure 1 shows a normative model for left mediodorsal nucleus as a function of age with the different quantiles. Ridge plots of z-score distributions showed shifts in specific nuclei associated with the conditions. For example, significant (Bonferroni p<0.05) shifts in z-score distributions showed a gradual progression from predominantly left anteroventral, mediodorsal, and pulvinar in early MCI to nearly all nuclei bilaterally in AD. Figure 2 summarizes extreme infranormal z-score deviations (<2) as a function of disease diagnosis. Note the progression from EMCI to AD as well as significant deviations in non-affective psychosis and schizophrenia. Supranormal z-score deviations (>2) were all non-significant.
