Participants
AD and aMCI subjects were recruited in this cross-sectional study conducted at the memory clinic of Zhejiang Provincial People’s Hospital(Hangzhou, China) from September 2016 to March 2019. The normal control (NC) subjects were volunteers recruited at the Hospital Health Promotion Center. All participants were right-handed and signed an informed consent. This study was carried out in accordance with the Declaration of Helsinki, and all procedures were approved by the local ethics committee of Zhejiang Provincial People’s Hospital (No. 2012KY002).
The medical history, neuropsychological test, physical examination, laboratory inspection, and craniocerebral MRI scan data of all subjects were collected. The neuropsychological scales involved the mini-mental state examination (MMSE) and Montreal cognitive assessment (MoCA).
The inclusion criteria for AD patients were: patients with AD met the revised NINCDS-ADRDA (National Institute of Neurological and Communicative Disorders and Stroke and the Alzheimer’s Disease and Related Disorders Association) criteria for “probable AD” with MMSE score ≤ 24 and MoCA score ≤ 26; higher scores indicated better cognition [22]. The inclusion criteria for aMCI patients were as follows: (1) complaints of memory impairment by patient, family, or physician; (2) the clinical manifestations were normal; (3) MMSE score > 24 and ≤ 27. The inclusion criteria for NC subjects were as follows: (1) no stroke, epilepsy, depression, or other neurological or mental diseases; (2) no hearing or visual impairment; (3) conventional craniocerebral MRI showed no infarction, hemorrhage, tumor, or other lesions; (4) MMSE score ≥ 28. The exclusion criteria for all three groups were: (1) stroke; (2) traumatic brain injury; (3) epilepsy, Parkinson’s disease, brain tumor, and other neurological diseases that led to memory impairment; (4) vascular dementia or other mixed dementia; (5) severe anemia, hypertension, diabetes, and use of psychotropic drugs. Finally, 89 AD patients, 24 aMCI patients, and 32 NCs were enrolled in this study.
Data acquisition
The MRI data were acquired on a clinical MR scanner using an eight-element receiving coil (Discovery MR750 3.0T; GE Healthcare, Wisconsin, USA). The MRI protocols included T1WI, T2WI, T2FLAIR, and diffusion-weighted imaging (DWI, b = 0 and 1000) to exclude subjects with craniocerebral disease. The DTI data were acquired using gradient echo single-shot EPI sequence with 25 diffusion gradient square [echo time (TE) = 63.8 ms, repetition time (TR) = 8612 ms, matrix = 256 × 256, field of view (FOV) = 192 × 192 mm2, slice thickness / slice spacing = 1.5 / 0 mm, 81 axial slices, voxel = 0.75 × 0.75 × 1.5 mm3, and b values 0 and 1000 s/mm2]. Also, high-resolution three-dimensional (3D) T1-weighted magnetization-prepared fast gradient echo (MPRAGE) vector images (TE = 2.9 ms, TR = 6.7 ms, matrix = 256 × 256, FOV = 256 × 256 mm2, slice thickness / slice spacing = 1 / 0 mm, 192 sagittal slices, and voxel = 1 × 1 × 1 mm3) and T2WI (TE = 0.1 ms, TR = 5.7 ms, matrix = 448 × 512, FOV = 448 × 512 mm2, slice thickness / slice spacing = 5.5 / 7 mm, 20 axial slices, and voxel = 1 × 1 × 5.5 mm3) were acquired.
DTI image preprocessing
Diffusion-weighted images were corrected for Eddy current distortions and gradient direction using FSL 6.0 [23]. High-resolution 3D T1WI and DTI brain images extraction were captured using CAT12 software [24]. Whole-brain volume of each subject was calculated by CAT12 for further statistical analyses. Then, the diffusion tensor model was applied at each voxel using the DTIfit program, and values of parameters such as fractional anisotropy (FA) and mean diffusivity (MD) were calculated in FSL; subsequently, the FA map was constructed.
Furthermore, we first used the linear registration command of FSL to register the 3D T1 image on the DTI individual space to achieve spatial consistency. Then, DTI images were registered on the MNI template by linear and nonlinear registration command to obtain the spatial transformation matrix file.
ALPS processing
After processing, we calculated the brain tensor in the x (left-right direction), y (anterior-posterior direction), and z (up-down direction) directions in each voxel using FSL by DTIfit. Then, four 5-mm-diameter spherical region of interest (ROI) was placed in the MNI T1 template (MNI152_T1_1mm_brain, 2.4) [25] (Fig. 1). The center coordinates of the left and right ROIs were (24, − 12, 24) and (− 28, − 12, 24) in projection fibers, and the centers of the left and right ROIs were (36, − 12, 24) and (− 40, − 12, 24) in association fibers, respectively [26]. The ROI masks were reversely registered to FA images of each subject using the deformation field generated in the above registration process. Manual verification confirmed the accuracy of registration and the location of ROIs for each subject in FSLEYES by clinical expert. Finally, the ALPS index was calculated as [(Dxxproj + Dxxassoc) / (Dyyproj + Dzzassoc)] on FA images (including left and right) [27] for each subject.
EPVS processing
The image was processed on the uAI research portal (uRP) [28], which integrates the general segmentation VB-Net for multiple ROIs [29]. In this study, we utilized the VB-Net for efficient and reproducible segmentation of EPVS.
The preprocessing steps for T1- and T2-weighted images were as follows: a) The N4 algorithm was applied to T1WI for bias correction to handle the inhomogeneity of the magnetic field; b) The skulls in T1WI were removed using the 3D VB-Net model deployed on the uRP; c) Advanced normalization tools (ANT) were used for the registration of T1- and T2-weighted images, and the brain mask of the T1WI was registered to T2WI, such that the skull of T2WI could also be removed; d) The voxel size of the T2-weighted images was resampled to 1 mm × 1 mm × 1 mm; e) The voxel intensities of T1 and T2-weighted images were rescaled to the range of [-1, 1].
The preprocessed T1WI were used as inputs to the 3D VB-Net model to segment the centrum semiovale, basal ganglia, and midbrain. These ROIs were used to localize EPVS lesions, and the preprocessed T1- and T2-weighted images were input into a 2D VB-Net model for PVS segmentation. Finally, PVS lesions < 2 mm in length were excluded as false positives, and the remaining PVS lesions were considered EPVS (Fig. 2).
To explore the segmentation performance of VB-Net, the Dice similarity coefficient (DSC) was calculated to evaluate the concordance level between the AI and the manual contouring results (as the ground truth). Notably, two clinical experts (Reader 1 and Reader 2) performed manual contouring of the EPVS. Four comparisons were made between the AI-segmentation results and ground truths (i.e., Reader 1, Reader 2, Readers’ union, and Readers’ intersection). The DSC was calculated as follows:
$$\text{D}\text{S}\text{C}=\frac{2\times TP}{2\times TP+FP+FN}$$
Where the TP denoted true positive, FP denoted false positive, and FN denoted false negative; the higher the DSC, the better the agreement with the ground truth. Additionally, recall and precision were computed to assess the segmentation performance of VB-Net using the following two equations.
$$\text{r}\text{e}\text{c}\text{a}\text{l}\text{l}=\frac{TP}{TP+FN}$$
$$\text{p}\text{r}\text{e}\text{c}\text{i}\text{s}\text{i}\text{o}\text{n}=\frac{TP}{TP+FP}$$
Based on the segmentation results, the total number and volume of EPVS lesions were output as PVS burden, and the BG, CSO EPVS volume fraction were used for subsequent experiments to further evaluate the characteristics of EPVS in specific brain regions, including comparison and correlation analyses. The calculation method involves dividing the volumes of EPVS in the BG and CSO regions by the total brain volume.
Receiver operating characteristic (ROC) curve of glymphatic system imaging markers
We used ROC curves to evaluate the predictive value of DTI-ALPS index, EPVS burden, and BG and CSO volume fraction as combined imaging markers for differentiating the three groups in AD-NC, AD-MCI, and MCI-NC. Also, the area under the curve (AUC) and the sensitivity and specificity were calculated.
Statistical analyses
SPSS 26.0 and FSL were used for statistical analysis, and the threshold was set at P < 0.05. First, SPSS 26.0 was used for statistical analysis of demographic and neuropsychological scales, and Shapiro–Wilk test was used to assess the normality of all data. Chi-square test was used to compare gender differences between groups.
Analysis of variance (ANOVA) or Kruskal–Wallis H tests were employed in accordance with the distribution of ALPS index and PVS indexes (EPVS burden, BG volume fraction) for statistical analysis. Due to the normal distribution of ALPS index results, we conducted ANOVA using a general linear model to compare the differences in ALPS index (left, right, and average value) among the AD, aMCI, and NC groups, including age, gender, education years, and whole-brain volume as covariates to obtain rigorous results. For positive results, multiple comparison analyses were corrected by Bonferroni’s or Dunn’s correction. Spearman’s correlation analysis was performed between ALPS index, EPVS burden, BG volume fraction, and cognitive scales. Furthermore, we conducted a correlation analysis between ALPS index result and PVS indexes to investigate the potential correlations between these two distinct dimensions of glymphatic neuroimaging markers in the disease (P < 0.05 was considered statistically significant).