This study used public data from the OASIS-3 dataset (https://central.xnat.org), including 43 MCI, 71 VMCI, and 87 matched HC. Briefly, OASIS-3 is a compilation of MRI and PET imaging data collected from several studies conducted by the Knight AD Research Center at the University of Washington over the past 15 years. The Clinical Dementia Rating (CDR) scale was used to assess the dementia status of uniform data set (UDS) (John C. Morris 2006). According to CDR, all participants were divided into different groups. Specifically, CDR = 0, 0.5 and 1 indicated HC, VMCI and MCI, respectively. Moreover, since the median Mini Mental State Examination (MMSE) could comprehensively and simply reflect the subjects’ mental status and degree of cognitive impairment, the MMSE score was collected for each subject (Tom N. Tombaugh 1992). All subjects had provided informed consent before MRI or neurological assessment. In addition, clinical scale information for all patients have been obtained. More detailed information is shown in Table 1.
The MRI images from all the participants were obtained using the 3-T Siemens’s Trio Tim scanners. All subjects were instructed to lie quietly and close their eyes during the scan. Resting state functional images were collected using an echo-planner imaging sequence with the following parameters: repetition time (TR) = 2200 ms, echo time (TE) = 27 ms, flip angle (FA) = 90°, number of slices = 33, slice thickness = 4 mm, voxel size = 4×4×4 mm3. For the T1-weighted images, the parameters are described as follows: voxel size = 1×1×1 mm3, echo time (TE) = 316 ms, repetition time (TR) = 2400 ms, flip angle = 8°, slice thickness = 1 mm. The echo plane imaging sequence was used to obtain diffusion tensor images (DTI) covering the entire brain, including 24 volumes with diffusion gradients applied along 24 non-collinear directions. The parameters of DTI are as follows: voxel size = 2×2×2 mm3, echo time = 0.112 s, repetition time = 14.5 s, flip angle = 90°, slice thickness = 2 mm.
Data preprocessing
Resting-state functional images and T1-weighted images preprocessing were performed by using SPM12 (http://www.fil.ion.ucl.ac.uk/spm/software/spm12) and Data Processing Assistant for Resting-State fMRI (http://rfmri.org/DPARSF). Briefly, the functional imaging preprocessing procedures consisted of the following: (1) To remove the unstable signal of the magnetic resonance scanner at the beginning of the scan, the first 5 time points were removed; (2) Head motion correction using rigid body translation and rotation, subjects with maximum motion > 3 mm or 3° were excluded; (3) Anatomical images were co-registered to the mean functional image using a trilinear interpolation with degrees of freedom; (4) For the T1-weighted image data, the DARTEL algorithm was used to segment the GM, WM and cerebrospinal fluid; (5) Regressing interference signal, including 24 head movement parameters and averaging cerebrospinal fluid signal. We did not perform regression on the global signal and the WM signal, to retain as much signals of interest as possible. We used scrubbing when observing movement "spikes" (frame displacement (FD) > 1mm), and performed a separate stopper to reduce movement effects; (6) Removed linear trends to correct signal drift; (7) To minimize the impact of non-neuronal signals on BOLD fluctuations, a band-pass filter of 0.01-0.1 Hz was used to extract the low-frequency components of functional images; (8) To avoid the confusion of WM and GM signals, the WM and GM templates were respectively used to minimize spatial smoothing of the functional images for each subject (4 mm full-width half-maximum [FWHM], isotropic); (9) The smoothed functional images were normalized from native space to MNI space with voxel size 3Í3Í3 mm³.
DTI were preprocessed and analyzed using FSL (http://www.fmrib.ox.ac.uk/fsl). For each subject, the preprocessing mainly includes removal of non-brain tissue (fractional intensity threshold was 0.2), correction of eddy current distortion, and local fitting of diffusion tensor (Yamada et al. 2014). MD coefficients were calculated based on voxel estimates, and the corresponding files were saved. Then, we performed Bayesian estimation of diffusion parameters obtained using sampling techniques (BEDPOSTX) processing. In this step, BEDPOSTX executes Markov Chain Monte Carlo sampling to establish the distribution of dispersion parameters on each voxel and performs Bayesian estimation at the same time. In order to match the information of the subjects to the same space for comparison, FMRIB’s linear image registration tool (FLIRT) was used to run standardization.
The hippocampus masks
To obtain the group-level hippocampus template, we adopted the FMRIB’s Integrated Registration and Segmentation Tool on induvidual structural images and obtained the hippocampus mask for each subject (Patenaude et al. 2011). The group-level hippocampus mask was obtained by averaging the individual hippocampus masks across all subjects. Subsequently, a strict threshold of 0.9 was selected to limit above hippocampus mask to obtain the final binarized group-level hippocampus mask. HIP.L and HIP.R were analyzed based on their functional anatomy and potential lateralization (C. Akos Szabo 2001); C Geroldi and colleagues' research based on nuclear magnetic resonance have shown that the bilateral hippocampus of normal adults was a reliable asymmetric structure and dementia was related to the change of normal anatomy asymmetry (C Geroldi 2000).
Creation of group-level WM and GM masks
To avoid mixing the WM and GM signals, we created the group-level WM and GM masks. Specifically, using the WM and GM images segmented from the above T1-weighted structure image, each voxel in the brain was identified with the maximum probability as WM or GM, which created a binary WM and GM mask for each subject. Then, binarized WM masks were averaged and then a threshold with 60% of subjects was used to create a binary group-level WM mask (Jiang et al. 2019a; Peer et al. 2017; Wang et al. 2020a). Adopting the same method, the binarized GM group-level mask was obtained, but using a lenient threshold with 20% of subjects. To further limit the WM and GM group-level masks, we compared the resulting masks to the functional images, and removed voxels identified as WM or GM yet having functional images in less than 80% of the subjects. Finally, to exclude the effect from deep brain structures, we identified the thalamus, caudate, nucleus putamen, globus pallidus and nucleus accumbens based on the Harvard-Oxford template and removed them from the group-level WM mask.
SFC maps with hippocampus as seed point
The current study explored the abnormal SFC between hippocampus and whole-brain voxels in VMCI and MCI subjects. To this end, the following steps were performed: (1) the averaged time series of HIP.L were extracted for each subject; (2) SFC was computed between above time series and all voxels time series within whole brain; (3) Fisher’s z transformation was performed for all correlation coefficients. Moreover, we also calculated the SFC between HIP.R and all voxels within whole brain using the above same steps.
Probabilistic tracking analysis
The probabilistic tracking analysis of diffusion tensor imaging (DTI) data was performed using FSL_6.0.3 (https://fsl.fmrib.ox.ac.uk/fsl/fslwiki/). We performed non-invasive probabilistic fiber bundle imaging using probtrackx2. In this step, the distribution estimated by BEDPOSTX was used for simulation. Before the fibers connecting the hippocampus and abnormal WM regions were tracked, these ROIs were transformed from MNI space to individual diffusion space by using FLIRT. Several regions that showed abnormal WM-SFC in patients were selected as regions of interest (ROIs) for analysis of DTI data, and were dilated by two voxels. Finally, FSL repeatedly samples from the main dispersion direction, calculates streamlines through these sampling points, and generates a set of probability streamlines, thereby extracting the fibers between the hippocampus and the abnormal ROI. Through multiple sampling, the prior distribution information was established, and then the true fiber distribution could be inferred from the prior information. The default 0.5 voxel step size, 5000 samples, and 2000 step size were used (–step length 0.5 -P 5000 -S 2000). Finally, we calculated the averaged MD of fiber bundles connecting the hippocampus and ROIs. Two-sample t-test was performed to explore the abnormal structural connectivity between the hippocampus and abnormal WM area between HC, VMCI and MCI (p < 0.05/numbers of ROIs, Bonferroni correction was used).
Statistical analysis
Within-group SFC between the HIP.L/HIP.R and voxels within the group-level GM mask was calculated by using one-sample t-test. Abnormal regions of SFC within group-level GM mask among three groups was identified by using one-way ANOVA, with age, gender, and education as covariates. Gaussian Random Field (GRF) theory was performed to correct for cluster-level multiple comparisons (minimum z scores > 2.3; cluster significance: p < 0.05, GRF corrected). Three abnormal ROIs were obtained for post-hoc analysis and were compared using the two-sample t-test with age, gender, and education as covariates (two-tailed, p < 0.05, Bonferroni-corrected for multiple comparisons (p < 0.05/3)).
Moreover, one-sample t-test was calculated for individual SFC maps based on the HIP.L/HIP.R as ROIs across participants in each of the three samples within group-level WM mask. The abnormal regions of SFC within group-level WM mask were obtained among three groups using above similar statistical analysis. Four ROIs were obtained for post-hoc analysis and were compared using the two-sample t-test with age, gender and education as covariates (two-tailed, Bonferroni-corrected for multiple comparisons, p < 0.05/4). Finally, Pearson correlation analysis was performed to explore the potential relationships between SFC of abnormal areas and clinical measures.