Data used in this study was obtained from the Alzheimer’s disease Neuroimaging Initiative (ADNI) database (adni.loni.usc.edu). The ADNI was launched in 2003 by the National Institute on Aging (NIA), the National Institute of Biomedical Imaging and Bioengineering (NIBIB), the Food and Drug Administration (FDA), private pharmaceutical companies and non-profit organizations, as a $60 million, 5-year public-private partnership. 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). Determination of sensitive and specific markers of very early AD progression is intended to aid researchers and clinicians in developing new treatments and monitor their effectiveness, as well as lessen the time and cost of clinical trials.

All procedures conducted in our study involving human participants were following the ethical rules of the institutional and national research committee and with the 1975 Helsinki declaration and its later amendments or comparable ethical standards. The ADNI project was authorized by the Institutional Review Boards of all participating institution and informed written consent was obtained from participants at each site.

We searched the potential subjects from the ADNI GO/ADNI 2 database in 2018, November, and firstly found 907 subjects (Figure 1. Flowchart of inclusion). Subsequently, we excluded the subjects as following steps (Braak et al., 2006) without CSF data to determine the biological profile; (Kalaria, 2016) with the suspected non-Alzheimer’s pathophysiology (SNAP) (Dani et al., 2017; Jack et al., 2018) without available T1- and T2-weighted MRI, and FDG PET to conducted imaging analyses; (Perrotin et al., 2015) without amyloid PET to complete biological profile double-confirmation; (Peng et al., 2015) with inconsistent pathological profile (i.e., discrepant amyloid status of amyloid PET and CSF); (Stricker et al., 2016) fail to data quality controls (e.g., excessive head motion, suffering specific symptom or diseases). Finally, we identified 252 subjects who undergone simultaneous scan (i.e., the time interval of scan less than 4 weeks) of MRI and FDG PET, lumbar puncture, and comprehensive neuropsychological assessments.

All participants underwent assessment of the general cognitive ability and disease severity (Mini-Mental State Examination, MMSE, and Clinical Dementia Rating, CDR, respectively), and other cognitive domains including memory (Wechsler Memory Scale LM and DM, immediate and delayed memory; Auditory Verbal Learning Test, AVLT, total and 30 min delayed score), processing speed (Trail-Making Test part A, TMT-A), visuospatial (Clock drawing Test, CDT), executive (TMT part B, TMT-B), and language ability (Category verbal fluency, CVF; Boston Naming Test, BNT).

Based on cognitive performance, we grouped subjects into cognitively unimpaired subjects, MCI, and dementia patients. Explicitly, we defined the cognitively unimpaired as subjects had CDR score of 0, MMSE between 24 and 30 (inclusive), DM ≥ 9 for subjects with ≥16 years of education; ≥5 for subjects with 8–15 years of education; and ≥3 for years of education ≤7; non-depression (Geriatric Depression Scale-15 score version <6) and non-dementia. We defined MCI as subjects preserved daily living activities, non-dementia, and objective cognitive impairment as shown on the DM and CDR score of 0.5. We defined dementia as: patients had MMSE score ≤ 26, CDR ≥ 0.5, and met the NINCDS/ADRDA criteria for probable AD.

Subsequently, we downloaded CSF dataset and identified the levels of Aβ1-42, total Tau (T-Tau) and phosphorylated tau (p-Tau181), which obtained using the standardized ADNI protocol as previously (Shaw et al., 2010). Based on previous ROC analyses of the autopsy-confirmed AD cases versus the NC group, we adopted the recommended cut-off of CSF biomarkers: 192 pg/ml for Aβ_1–42 and 23 pg/ml for p-Tau₁₈₁ (De Meyer et al., 2010; Shaw et al., 2010). Although the NIA-AA research framework suggests that either CSF or amyloid PET enables to determine one’s biological AD profile, in some cases, CSF Aβ_1–42 might become abnormal before amyloid PET (Bateman et al., 2012; Ben Bouallegue et al., 2017). Based on these evidence, some previous studies suggested that the combination of amyloid CSF and PET biomarker may further improve the accuracy of biological classification to some extent (Bouter et al., 2019; Spallazzi et al., 2019). Thus, we adopted the recommended composite Florbetapir cutoff value of 1.11 (whole cerebellum as reference region) (i.e., A+ represents amyloid PET SUVR composite score > 1.11), which is equivalent to the upper 95% confidence interval above the mean of a group of young healthy controls (Landau et al., 2013). Finally, we excluded 11 subjects with the inconsistency results of amyloid status. We also excluded 74 subjects with normal Aβ_1–42 and abnormal p-Tau₁₈₁ (A−T+) since it might represent the non-AD related pathology.

Finally, we classified all subjects into four groups according to their cognitive performance and biological profile (Jack et al., 2018). Exactly, (a) controls: consisting of cognitively unimpaired subjects with normal Aβ_1–42, p-Tau₁₈₁ and negative amyloid status based on PET (A-T-); (b) preclinical AD, consisting of cognitively unimpaired subjects with abnormal Aβ_1–42, normal p-Tau₁₈₁, and positive amyloid status based on PET (A+T+); (c) prodromal AD, consisting of MCI with abnormal Aβ_1–42, p-Tau₁₈₁ and positive amyloid status based on PET (A+T+); (d) AD with dementia, consisting of demented subjects with abnormal Aβ_1–42, P-Tau₁₈₁ and positive amyloid status based on PET (A+T+).

Apolipoprotein (APOE) genotyping for each subject was performed as previously described (Saykin et al., 2010). In brief, genotyping of all subjects for APOE allele status was analyzed using DNA extracted from peripheral blood cells. The cells were collected in single EDTA plastic tubes (10 ml) and were sent via overnight delivery, at room temperature, to the University of Pennsylvania AD Biofluid Bank Laboratory.

Details of the ADNI neuroimaging acquisition protocol are publicly available on the Laboratory of Neuroimaging (LONI) website. Structural images were acquired based on 3D Magnetization Prepared Rapid Acquisition Gradient Echo (MPRAGE) T1W sequence and fast spin-echo/turbo spin-echo (FSE/TSE) sequence. T1W image parameters: voxel size = 1.1 × 1.1 × 1.2 mm3; echo time (TE) = 2.98 ms; repetition time (TR) = 2300 ms. T2W image parameters: TR = 4000 ms; TE = 78 ms; voxel size = 0.9375 × 0.9375 × 4 mm3. The FLAIR images were obtained using an echo-planar imaging sequence with the following parameters: TR = 9000 ms, TE = 90 ms, and TI = 2500 ms. Meanwhile, FDG PET images were acquired by either a 30-min six frame scan acquired 30 to 60-min post-injection or a static 30-min single-frame scan acquired 30 to 60-min post-injection (Landau et al., 2011). Amyloid PET data were acquired during 4 × 5min time frames measured 50–70 min post-injection of ¹⁸F-AV45. More detailed information can be found online regarding PET scanning protocol: http://adni.loni.usc.edu/wp-content/uploads/2010/05/ADNI2_PET_Tech_Manual_0142011.pdf.

We processed the T1W images using the Computational Anatomy Toolbox (CAT, version 12¹) in SPM12. Concisely, we firstly aligned the T1W images to the standardized template image. We segmented the standardized T1W images into the compartments of CSF, WM, and GM with the bias correction. Sequentially, we normalized the GM density maps to the MNI space as 1 × 1 × 1 mm³ voxels and smoothed using the 8 mm FWHM Gaussian kernel. Finally, we estimated the modulated GM maps.

The T1W/T2W ratio mapping was processed on the basis of the workflow, MRtool² implemented in the SPM12 software, described by Ganzetti et al. (2014). This standardized pipeline comprises the bias correction and intensity calibration on both the T1- and T2 weighted images and the subsequent computation of the ratio value between the pre-processed T1-weighted and T2-weighted images.

Specifically, the original T2W image was firstly co-registered to the T1W image through a rigid-body transformation. Next, we separately the subjected T1W and T2W images to bias correction. After correcting for the intensity non-uniformity, the T1W and T2W images were further processed to standardize their intensity by using a linear scaling procedure. After calibrating the T1W and T2W images, their ratio was calculated to generate the calibrated ratio map. To conduct two-sample T-test comparisons between groups, we spatially transformed the calibrated ratio map from the individual subject space to the Montreal Neurological Institute (MNI) space. According to the research hypothesis, we extracted the GM components of the T1w/T2w ratio maps based on the segmented normalized T1W image. Before the statistical analyses, we smoothed the T1W/T2W ratio map with a Gaussian kernel of 6 mm full-width at half maximum (FWMH).

We downloaded the ¹⁸F FDG PET and amyloid PET data from the ADNI database in its most processed format (PET Pre-processing protocol³, University of Berkeley). Specifically, these pre-processed scans were created following co-registration dynamic (to the first frame), averaging frames, spatial re-orientation (AC-PC line), intensity normalization (within subject-specific mask), and smoothing (uniform isotropic spatial resolution 8 mm FWMH). To exclude the between-subject nuisance variability in tracer uptake, we normalized the intensity of FDG PET by dividing it by the cerebellar pons/vermis; also, we normalized the intensity of amyloid PET by dividing it by the whole cerebellum. Next, the produced standardized uptake values (SUVR) maps entered the followed statistical analyses (Landau et al., 2011).

Given that the white matter hyperintensities (WMH) may correlate with the demyelination and metabolism abnormalities, we compared the WMH load difference between groups in the current study (Jiaerken et al., 2018). For each subject, their 3D T1-weighted scans and FLAIR scans were normalized to MNI space. Then the Lesion Segmentation Toolbox (LST) in SPM12 was used to produce the WMH lesion segmentation maps based on the 3D T1-weighted scans and FLAIR scans (Schmidt et al., 2012). Automatically generated lesion maps were further manually reviewed for misclassification by two experienced neuroradiologists. Next, the WMH volume for the each subject was calculated by multiplying the voxel number by 1 mm³. More information can be obtained in our previous work (Luo et al., 2017, 2018).

We conducted the statistical analyses utilizing IBM SPSS 24 statistical software for Windows. Group differences in the age, education level, neuropsychological scores, and CSF were tested by the ANOVA, followed by post hoc two-sample T-tests if ANOVA was significant (p < 0.05, Bonferroni corrected).

We adopted the two-sample T-tests to examine the difference between controls and patient groups in the intracortical T1W/T2W ratio mapping, VBM, and FDG PET data. Notably, all imaging analyses were restrained within a standardized group gray matter (GM) mask, created from the CAT12 and SPM Masking Toolbox (Ridgway et al., 2009). In a voxel-wise way, we performed the imaging statistical in the DPABI, which is based on SPM 12 (Yan et al., 2016). To correct the multiple comparisons, we utilized the Gaussian random field (GRF) theory planted in the statistical module of the DPABI (Yan et al., 2016). Firstly, the p maps (statistical map) were converted to the Z maps according to the normal inverse cumulative distribution function, with the sign of group mean differences applied. The thresholding approach we used is single-voxel thresholds (cluster-defining thresholds) of p < 0.005 (Z > 2.58), and cluster size thresholds of p < 0.05 (Altman and Bland, 1994; Nichols and Hayasaka, 2003; Chen et al., 2018). Given the possible effects of age, gender, and APOE ε4 status, we also controlled these factors as covariates in the following analyses. Furthermore, given that the possible impacts of GM atrophy on T1W/T2W ratio, we also measured the group difference after controlling the GM segmented map in a voxel-wise way. We conducted these statistical analyses based on DPABI (Yan et al., 2016; Chen et al., 2018).

To investigate the possible pathological mechanism of imaging metrics, we performed partial correlation analyses among imaging metrics, neuropathological, and neuropsychological data. Specifically, we obtained the mean value of T1W/T2W ratio and FDG PET SUVR from regions showing significant between-group difference below. Based on SPSS, we performed a series of correlation analyses used age, gender, and APOE ε4 status as covariates and Bonferroni correction approach. The variables in correlation analyses included the mean value of T1W/T2W ratio, FDG PET SUVR, amyloid PET SUVR in bilateral inferior parietal lobe and hippocampus regions, CSF data (Aβ_1–42, T-Tau, p-Tau₁₈₁), general and multi-domain cognitive abilities score. After multiple comparison correction, some correlation relationship was no longer significant, we still showed these results in the manuscript and labeled them as “uncorrected” for the explorative purpose.

We presented the quantitative variables as the mean and standard deviations (SD), and the categorical variable as absolute and relative frequency. There were no statistical differences in term of education level, gender composition, and white matter hyperintensities among four groups. Concerning the pair-wise comparison, preclinical AD only had inferior AVLT performance than controls. Table 1 summarizes the demographics, CSF biomarker levels, and neuropsychological evaluation of the subjects (Supplementary Materials 1–3 shows post hoc two-sample T-test).

As expected, we observed that cognitively unimpaired preclinical AD subjects had lower T1W/T2W ratio value in the right inferior parietal lobule (IPL) than controls (Figure 2, top panel). As the disease progresses, prodromal AD subjects had a lower ratio value in two-sided IPL regions than controls. Ultimately, AD dementia had the most widespread T1W/T2W ratio decrease, involving the bilateral hippocampus (HP) and bilateral IPL, relative to controls (Figure 3). Controlling the GM volume in a voxel-wise way, no T1W/T2W ratio value differences existed between controls and cognitively unimpaired preclinical AD (Figure 2, bottom panel). But the prodromal AD subjects had decreased ratio value in the bilateral IP and left middle temporal gyrus compared to controls. Concerning AD dementia, the results remained almost the same. After taking the age, gender, and APOE ε4 status into consideration, significant difference regions become narrow, the ratio value difference between controls and preclinical AD no longer existed, and the right-side IPL differential region was no longer in existence. Supplementary Material 12 illustrates the imaging results corrected by age, gender and APOE ε4 status, and Supplementary Materials 4–7 shows the MNI coordinate information.

The Figure 2 (second row) shows the GM volume group-difference maps. We found no gray matter volume differences between controls and preclinical AD. As the disease progresses, bilateral HP atrophy can be found in prodromal AD subjects, when compared to controls. At the stage of AD dementia, subjects had widespread GM atrophy relative to controls, particularly in the bilateral medial temporal lobe. After taking the age, gender, and APOE ε4 status into consideration, results remained almost unaltered (Supplementary Materials 8, 9).

Neither preclinical AD nor prodromal AD had decreased metabolism relative to controls (Figure 2). Only the AD dementia patients had increased metabolism in the left middle cingulate cortex, and widespread metabolism decrease compared to controls, especially in the bilateral middle temporal gyrus (MTG). After taking the age, gender, and APOE ε4 status into consideration, regions showing increased metabolism was no longer in existence (Supplementary Materials 10, 11).

We showed the results centering on the T1W/T2W ratio and put completed correlation results in Supplementary Material 13. Within patient groups (A+T+), the left (r = −0.15) and right IPL (r = −0.17) T1W/T2W ratio value related to the T-tau level (p < 0.05 uncorrected), the left HP T1W/T2W ratio value related to the Aβ_1–42 level (r = 0.14, p < 0.05, uncorrected). Within the patient groups, we also found that T1W/T2W ratio related to amyloid SUVR in the left (r = 0.41, p < 0.05 corrected) and right HP (r = 0.43, p < 0.05 corrected). However, no correlation relationship between the T1W/T2W ratio and the amyloid SUVR in IPL existed (Figure 4).

As for the clinical correlation relationships, within patient groups, we found that T1W/T2W ratio, FDG PET and HP volume correlated with general cognitive and multiple cognitive domains. These results suggested that all three metrics were effective imaging biomarkers. Further, all three metrics correlated with each other. Especially, the T1W/T2W ratio related to the FDG SUVR in term of left (r = 0.40, p < 0.05 corrected) and right (r = 0.38, p < 0.001) HP.