2.1 Demographics
Data used in this study were obtained from the Alzheimer’s Disease Neuroimaging Initiative (ADNI) (adni.loni.usc.edu). The ADNI was launched in 2003 as a public-private partnership, led by Principal Investigator Michael W. Weiner, MD. The primary goal of ADNI has been to test whether serial magnetic resonance imaging (MRI), positron emission tomography (PET), other biological markers, and clinical and neuropsychological assessment can be combined to measure the progression of mild cognitive impairment (MCI) and early Alzheimer’s disease (AD). For up-to-date information, see www.adni-info.org.
From the ADNI database, we analyzed subjects who underwent both MRI and PET (amyloid, AV 45 and tau, AV 1451) including 78 cognitively normal (CN), 50 EMCI, 34 LMCI, and 39 AD individuals. Subjects were sampled according to the following criteria: age, around 60 to 90 years old, education, 12 to 20 years, and gender-matched within groups. To assess the AD continuum, amyloid-negative CN and amyloid-positive EMCI, LMCI, and AD subjects were selected. The EMCI group was subdivided into 38 dementia non-converters (stable EMCI) and 12 converters to assess changes in disease progression. A total of 201 subjects’ T1 and DTI images were gathered from the ADNI. To increase the sample size, a multi-center approach was used, as discussed in [13]. The amyloid positivity of subjects was determined using whole brain PET AV45 standardized uptake value ratio (SUVR) with a 1.11 cutoff. Table 1 shows the demographics of the subjects used in this study; note that 44 CN subjects, nine EMCI subjects, five LMCI subjects, and three AD subjects underwent AV1451 tau PET imaging. An additional 28 CN subjects with amyloid positivity were analyzed to identify the earliest AD pathological changes as presented in Supp. Table 1.
2.2 Image processing
T1-weighted images were processed using FreeSurfer package v6.0 (http://surfer.nmr.mgh.harvard.edu) as previously reported in [13]. Cortical thickness (CTh) maps were registered to the FreeSurfer average sphere using spherical registration for group comparison. DTI and PET images were registered with respect to T1 images using a boundary-based algorithm for further processing. DTI images were processed using the FSL package as follows: eddy current correction, rotate gradient vectors from the results of eddy correction, and tensor fitting to produce the MD map and primary eigenvector map. DTI metrics were further processed to avoid partial volume effects following Koo et al. [27]. PET images were partial volume corrected using mri_gtmpvc which is built into the FreeSurfer package. PET images were normalized by mean signal from the whole cerebellum and converted to SUVR for amyloid and tau PET, AV45, and AV1451, respectively. The images were then boundary-based registered to corresponding T1 structural images. To avoid any partial volume effects, the center parts of the cortical column were sampled for surface analysis. Lastly, CTh was smoothed with a 10-mm full width half maximum Gaussian kernel, while other modalities were smoothed with a 15-mm kernel. Fig. 1 shows an overall schematic of the process.
2.3 Calculation of radiality
A surface normal vector was obtained from the individual gray matter surface to define the cortical orientation. FreeSurfer represents the surface in triangular meshes, and the surface normal vector can be computed using the cross-product between edges. The vertex-wise dot product between the primary diffusion direction, primary eigenvector of the diffusion tensor, and the surface normal vector was quantified as a radiality index, r, where v represents the surface normal vector and represents the primary diffusion direction [22].
See formula 1 in the supplementary files.
It ranges from 0 to 1, where r = 0 indicates tangential diffusion and r = 1 indicates radial diffusion to the cortex. The subject’s principal eigenvector map was projected onto the individual surface reconstruction to calculate vertex-wise radiality, as discussed in [22].
2.4 Cutoff analysis
To further test the feasibility of radiality as an AD biomarker, we performed cutoff analysis using receiver operating characteristic graphs to distinguish CN from different AD stages, as shown in Supp. Table 2. The feature used was the mean radiality within the cluster obtained from the CN vs. EMCI group comparison. By varying the cutoff, we sought to find the cost-effective point where it minimizes the difference between sensitivity and specificity [23].
2.5 Statistical analysis
We first compared the differences between groups for radiality, CTh, MD, AV45, and AV1451 using a general linear model, which is available in FreeSurfer. The results were cluster-wise corrected for a family-wise error (FWE)-corrected p-value < 0.05.
To test the associations between radiality and other neuroimaging biomarkers, we calculated a set of vertex-wise partial correlations with radiality as the dependent variable and CTh, MD, AV45, and AV1451 as the independent variable. Age, gender, years of education, and MRI center were set as covariates for cluster analyses. A permutation test was applied to account for multiple comparisons using a Monte Carlo simulation with 10,000 repeats, which is a built-in function of FreeSurfer.
To test the linear relationship between radiality and other neuroimaging biomarkers, we quantified mean metrics within AD-specific ROIs. ROIs include the entorhinal, fusiform, insula, inferior, middle, and superior temporal cortex. Mean metrics within ROIs were plotted in box and whisker plots and are presented in Fig. 5(e) and Fig. 6. Significant differences between groups were tested using one-way analysis of variance (ANOVA).