Diagnostic Efficacy of Voxel-Mirrored Homotopic Connectivity in Vascular Dementia as Compared to Alzheimer’s Related Neurodegenerative Diseases—A Resting State fMRI Study

Previous studies have demonstrated that functional connectivity (FC) of different brain regions in resting state function MRI were abnormal in patients suffering from mild cognitive impairment (MCI) and Alzheimer’s disease (AD) when comparing to healthy controls (HC) using seed based, independent component analysis (ICA) or small world network techniques. A new technique called voxel-mirrored homotopic connectivity (VMHC) was used in the current study to evaluate the value of interhemispheric functional connectivity (IFC) as a diagnostic tool to differentiate vascular dementia (VD) from other Alzheimer’s related neurodegenerative diseases. Eighty-three participants were recruited from the university hospital memory clinic. A multidisciplinary panel formed by a neuroradiologist and two geriatricians classified the participants into VD (13), AD (16), MCI (29), and HC (25) based on clinical history, Montreal Cognitive Assessment Hong Kong version (HK-MoCA) neuropsychological score, structural MRI, MR perfusion, and 18-F Flutametamol (amyloid) PET-CT findings of individual subjects. We adopted the calculation method used by Kelly et al. (2011) and Zuo et al. (2010) in obtaining VMHC maps. Specific patterns of VMHC maps were obtained for VD, AD, and MCI to HC comparison. VD showed significant reduction in VMHC in frontal orbital gyrus and gyrus rectus. Increased VMHC was observed in default mode network (DMN), executive control network (ECN), and the remaining salient network (SN) regions. AD showed a reduction of IFC in all DMN, ECN, and SN regions; whereas MCI showed VMHC reduction in vSN, and increased VMHC in DMN and ECN. When combining VMHC values of relevant brain regions, the accuracy was improved to 87%, 92%, and 83% for VD, AD, and MCI from HC, respectively, in receiver operating characteristic (ROC) analysis. Through studying the VMHC maps and using VMHC values in relevant brain regions, VMHC can be considered as a reliable diagnostic tool for VD, AD, and MCI from HC.


Clinical and Neuropsychological Assessment
Vascular risk factors (including hypertension, hyperlipidemia and diabetes) of all participants were obtained, either from an interview conducted by a nurse, or from medical record retrieved from the hospital clinical management system database. All participants underwent HK-MoCA which was administered by a trained research assistant [28].

Pre-Processing of Resting State Functional Images
Data pre-processing was conducted by Data Processing Assistant for Resting State fMRI (DPARSF) v.4.4 (http://www.rfmri.org/dparsf, accessed on 9 October 2021) with Statistical Parametric Mapping software (SPM12 http://www.fil.ion.ucl.ac.uk/ spm/software/spm12/, accessed on 9 October 2021) on a MATLAB platform 2018a (The MathWorks, Inc., Natick, MA, USA). Preprocessing was performed with standard processing steps suggested by the developer of DPARSF [29], with details as follows: the first 10 images volume from each subject was discarded to eliminate magnetic saturation Life 2021, 11, 1108 4 of 18 effects, and the remaining 170 images were re-sliced for time and motion correction (the 35th slice was used as reference). A maximum displacement in any direction (x, y, and z) of 2 mm and maximum rotation (x, y, and z) of 2 degrees during the rs-fMRI scan were limited. To spatially normalize the fMRI image, the high-resolution individual T1-weighted images were co-registered to the mean fMRI; the resulting aligned images were segmented and transformed into Montreal Neurological Institute (MNI) space using the DARTEL toolbox, and a group template was generated. Then, the motioncorrected functional images were specifically normalized to the group template using the transfer parameter estimated through DARTEL segmentation and resampled to voxel size of 3 × 3 × 3 mm 3 . The fMRI images were smoothed with Gaussian kernel of 4 × 4 × 4 mm 3 full width half maximum (FWHM). To minimize low-frequency drift and high-frequency physiological noise, temporal band-pass filtering (0.01-0.1 Hz) was applied. Frame-wise displacement (FD) threshold of 0.5 was applied to reduce the motion artefacts, where bad points were interpolated using the nearest neighbor algorithm. Finally, a nuisance linear regression was performed with the white matter, cerebrospinal fluid, global signal, six head motion parameters at one time point earlier, and the 12 corresponding squared items (Friston 24-parameter model) as covariates.

Voxel-Mirrored Homotopic Connectivity VMHC
For VMHC, pre-processed images were transformed to a symmetric brain template with the application of an rs-fMRI image. The template was obtained by flipping the left or right hemispheres along the X-axis midline and average with the original image. From each participant, the T1w and rs-fMRI images were normalized to the MNI space and co-registered to this group-specific symmetric template. The interhemispheric functional connectivity was analyzed based on the Pearson's correlation between paired voxels located in exactly the same location but in a different hemisphere. The resultant correlations composed the VMHC map, the VMHC value of a particular region of interest (ROI) could be extracted from the corresponding VMHC map for further comparison. All the results were presented using the unilateral (left side only) brain mask, due to the fact that VMHC values were symmetric. VMHC values of DMN, ECN, SN and supplementary motor area, a total of 36 regions, were recorded.

Leukoaraiosis and Brain Regional Volume Segmentation
Leukoaraiosis, also called white matter hyper-intensity, was described as subcortical hyper-intensity lesions in the white matter, either diffuse in periventricular regions, or multifocal [30]. It can be shown in T2W FLAIR image as hyper-intensity. Wei et al. (2019) showed that white matter hyper-intensities from T2-FLAIR is equivalent to the while matter hypo-intensities derived from T1W MRI by FreeSurfer [31]. In the current study, brain regional volumes and leukoaraiosis volume were calculated from T1W MRI using the FreeSurfer software package (http://surfer.nmr.mgh.harvard.edu/ ver7.1, accessed on 10 October 2021) [32]. The subcortical volume segmentation and measurement of different brain structure were done by automated procedures [33]. Briefly, the process contained motion correction, removal of non-brain tissue using a hybrid watershed/surface deformation procedure [34] automated Talairach transformation, segmentation of the subcortical white matter and deep gray matter volumetric structures (including the hippocampus, amygdala, caudate, putamen, and ventricles) [33,34], intensity normalization, tessellation of the gray-white matter boundary, automated topology correction [35,36], and surface deformation following intensity gradients to optimally place the gray-white matter and gray matter/CSF borders at the location where the greatest shift in intensity defines the transition to the other tissue class [37][38][39].

Statistical Analysis
Demographic data including age, gender, HK-MoCA scores between AD, VD, MCI, and HC groups were analyzed using one-way ANOVA on rank test with post-hoc Bonfer-Life 2021, 11, 1108 5 of 18 roni test. p < 0.05 was considered to be significantly different in all tests. Analyses were performed using SPSS software version 25.0 (SPSS Inc., Chicago, IL, USA). VMHC value between different groups was examined in a voxel-wise manner based on the DPARSF software. A two sample independent t-test was performed to compare the participant images of each group with HC to produce a t-map. Minimum cluster size was set to 100 voxels to minimize the effect from noise. The significance level of the group difference was threshold based on the two-tailed Gaussian Random Field (GRF) theory, with voxel level of p < 0.01 and cluster level of p < 0.05. A VMHC map was obtained. A location with peak t-value was recorded as region of interest (ROI). Brain regional volume and leukoaraiosis volume between different groups were examined using MANCOVA test, using age, gender, and total intracranial volume as covariates. To investigate whether the IFC was affected by the volume, we studied the relationship between the VMHC values and their corresponding brain regional volumes using partial correlation analysis, with age and sex as covariates.
To access the diagnostic efficacy of neuronal spontaneous activity, the VMHC value of brain regions derived based on AAL (anatomical automatic labeling) template were extracted using REST software version 1.8. The predictive performance of VMHC values was calculated using Statistical Product and Service Solutions (SPSS) Ver25.0 based on the results of specificity, sensitivity, and area under the receiver operating characteristics (ROC) curves (AUC). The AUC values were graded i.e., 0.7-0.8 as fair, 0.8-0.9 as good, and 0.9-1 as excellent [40]. Optimal threshold in every brain region was determined by maximizing the Youden's index (sensitivity + specificity −1) and accuracy. To combine the effect of altered VMHC brain regions, binary logistic regression analysis was conducted using SPSS version 25.0. The statistical significance was set as p < 0.05.

Demographics, Clinical and Neuropsychological Assessments
Findings on demographics, clinical and neuropsychological test scores (HK-MoCA) of VD, AD, MCI, and HC are listed in Table 1. The age of VD and MCI was significantly higher than that of HC with p < 0.05 and p < 0.01, respectively. The HK-MoCA scores in AD were significantly lower than that of the other groups, and all diseased groups were significantly lower than that of HC. The leukoaraiosis volume in VD was significantly larger than the other three groups. Additionally, a higher percentage of participants in the VD group had T2DM, hypertension and hyperlipidemia when compared to other groups.

VMHC Map in VD, AD and MCI Compared to HC
The two-sample independent t-test VMHC maps of diseased groups as compared to HC in both global and regional levels are presented in Figure 1, i.e., VD vs. HC, AD vs. HC, and MCI vs. HC. Compared to the HC, VD showed significant reduction in VMHC in superior and medial frontal orbital gyrus, gyrus rectus. Increased VMHC were observed in superior frontal gyrus, inferior frontal gyrus operculum and triangular, precentral gyrus, rolandic operculum, supplementary motor area, insula, anterior, middle and posterior cingulate, hippocampus, parahippocampus, amygdala and caudate when compared to HC (p < 0.05 GRF corrected), as shown in Figure 1a. AD showed significantly decreased VMHC in the DMN, ECN, and SN (p < 0.05 GRF corrected), as shown in Figure 1b. MCI showed significant VMHC reduction in superior, medial, and inferior frontal orbital gyrus, olfactory and gyrus rectus. Increased VMHC were observed in supplementary motor area, precentral gyrus, postcentral gyrus, rolandic operculum, insula, middle cingulate, paracentral lobule, caudate, putamen and thalamus when compared to HC (p < 0.05 GRF corrected), as shown in Figure 1c. orbital gyrus, olfactory and gyrus rectus. Increased VMHC were observed in supplementary motor area, precentral gyrus, postcentral gyrus, rolandic operculum, insula, middle cingulate, paracentral lobule, caudate, putamen and thalamus when compared to HC (p < 0.05 GRF corrected), as shown in Figure 1c.

VMHC Analysis
From the above maps, brain regions with significant different VMHC values (T value) between groups were obtained. The results are listed in Tables 2-4. Negative intensity indicates lower VMHC values of the diseased group when compared to HC. Negative volume indicates regional volume reduction when compared to HC.    Tables 2-4. Brain regions that showed significant differences in VMHC values when compared to HC. All cluster size >100 voxels, adjusted for age, gender, and total intracranial volume. # indicated significant volume change between the two groups.

Correlation between VMHC Values and Their Corresponding Brain Regional Volumes
Compared the cognitive impaired groups with HC, the regions of interest with significant volume reduction (p < 0.05) are listed in Tables 2-4. In VD, hippocampus and amygdala showed atrophy, but there was no significant correlation between VMHC values and their corresponding volumes. In AD, more regions showed atrophy when compared to VD, but only middle occipital gyrus showed a negative correlation r = 0.692 with VMHC values. In MCI, there was no significant correlation between VMHC values and their corresponding volumes. In conclusion, the variations in VMHC values were not related to the volume variations in all studied groups. ROC analysis demonstrated that the regional VMHC changes exhibit fair performance in distinguishing VD patients from HC. Reduced VMHC was found in calcarine (accuracy = 70%, p = 0.034), lingual gyrus (accuracy = 75%, p = 0.034), and gyrus rectus (accuracy = 78%, p = 0.026). Increased VMHC was found at Supplementary motor area (accuracy = 68%, p = 0.029). Combining the VMHC values of these four regions, the accuracy improved to 87% with AUC over 0.9. Balanced sensitivity and specificity were achieved. Details are listed in Table 5. The ROC curves are shown in Figures 2 and 3.

Correlation between VMHC Values and Their Corresponding Brain Regional Volumes
Compared the cognitive impaired groups with HC, the regions of interest with significant volume reduction (p < 0.05) are listed in Tables 2-4. In VD, hippocampus and amygdala showed atrophy, but there was no significant correlation between VMHC values and their corresponding volumes. In AD, more regions showed atrophy when compared to VD, but only middle occipital gyrus showed a negative correlation r = 0.692 with VMHC values. In MCI, there was no significant correlation between VMHC values and their corresponding volumes. In conclusion, the variations in VMHC values were not related to the volume variations in all studied groups.

ROC for VD vs. HC
ROC analysis demonstrated that the regional VMHC changes exhibit fair performance in distinguishing VD patients from HC. Reduced VMHC was found in calcarine (accuracy = 70%, p = 0.034), lingual gyrus (accuracy = 75%, p = 0.034), and gyrus rectus (accuracy = 78%, p = 0.026). Increased VMHC was found at Supplementary motor area (accuracy = 68%, p = 0.029). Combining the VMHC values of these four regions, the accuracy improved to 87% with AUC over 0.9. Balanced sensitivity and specificity were achieved. Details are listed in Table 5. The ROC curves are shown in Figures 2 and 3.

ROC for AD vs. HC
ROC analysis demonstrated that the regional VMHC changes of inferior parietal (accuracy = 81%, p = 0.001), superior temporal gyrus (accuracy = 80%, p = 0.001), and inferior occipital gyrus (accuracy = 80%, p = 0.01) exhibited good performance in distinguishing AD patients from HC. Balanced sensitivity and specificity were achieved in these three regions. The other 12 regions showed fair accuracy in the ROC analysis. After combining the VMHC values of all 15 regions, the accuracy improved to 92% with AUC over 0.9. Details are listed in Table 6. The ROC curves are shown in Figure 4.  ROC analysis demonstrated that the regional VMHC changes of inferior parietal (accuracy = 81%, p = 0.001), superior temporal gyrus (accuracy = 80%, p = 0.001), and inferior occipital gyrus (accuracy = 80%, p = 0.01) exhibited good performance in distinguishing AD patients from HC. Balanced sensitivity and specificity were achieved in these three regions. The other 12 regions showed fair accuracy in the ROC analysis. After combining the VMHC values of all 15 regions, the accuracy improved to 92% with AUC over 0.9. Details are listed in Table 6. The ROC curves are shown in Figure 4.  ROC analysis demonstrated that regional VMHC changes exhibit fair performance in distinguishing MCI from HC. Activated VMHC was found in inferior frontal operculum (accuracy = 70%, p = 0.031), Rolandic operculum (accuracy = 68%, p = 0.026), and supplementary motor area (accuracy = 68%, p = 0.006); attenuated VMHC was found in inferior frontal orbital gyrus (accuracy = 68%, p = 0.031), gyrus rectus (accuracy = 72%, p = 0.003), and superior frontal orbital gyrus (accuracy = 68%, p = 0.03). After combining the VMHC values of inferior frontal operculum, Rolandic operculum, supplementary motor area, inferior frontal orbital gyrus, and gyrus rectus, the accuracy improved to 83% with AUC over 0.9. Balanced sensitivity and specificity were achieved. Details are listed in Table 7. The ROC curves are shown in Figures 5 and 6.   ROC analysis demonstrated that regional VMHC changes exhibit fair performance in distinguishing MCI from HC. Activated VMHC was found in inferior frontal operculum (accuracy = 70%, p = 0.031), Rolandic operculum (accuracy = 68%, p = 0.026), and supplementary motor area (accuracy = 68%, p = 0.006); attenuated VMHC was found in inferior frontal orbital gyrus (accuracy = 68%, p = 0.031), gyrus rectus (accuracy = 72%, p = 0.003), and superior frontal orbital gyrus (accuracy = 68%, p = 0.03). After combining the VMHC values of inferior frontal operculum, Rolandic operculum, supplementary motor area, inferior frontal orbital gyrus, and gyrus rectus, the accuracy improved to 83% with AUC over 0.9. Balanced sensitivity and specificity were achieved. Details are listed in Table 7. The ROC curves are shown in Figures 5 and 6.

Discussion
Our study investigated rs-fMRI IFC in VD, AD, and MCI, and produced three main findings: first, VD, similar to AD and MCI, showed aberrant VMHC when compared to HC. Secondly, the aberrant VMHC was independent to its corresponding brain regional volume but may relate to the leukoaraiosis and vascular risk factors in VD, and related to amyloid burden in AD and MCI. Thirdly, combining regional VMHC achieved high accuracy to differentiate VD, AD, and MCI from HC, with balanced sensitivity and specificity in ROC analysis. Thus, VMHC has the potential to be a good diagnostic tool in characterizing the three cognitive impaired groups from HC.

VMHC and Brain Regional Volume Change Are Two Independent Metrics
Our results demonstrated that brain regional volume reduction was independent to the variation of VMHC values. The results were coherent with our expectations. VMHC measured IFC was affected in the early stage of dementia; whereas, structural changes would usually exist in the later stage of the disease. The relationship between VMHC and brain regional volume was weak and indirect.

Aberrant VMHC in the Cognitive Impaired Groups and Its Diagnostic Accuracy
Interhemispheric coordination impairment is a classical deficit of patients suffering from dementia. Compare to structural imaging, it is a sensitive biomarker which is useful in early diagnosis. To study interhemispheric coordination, EEG was used by Duffy and his team [41]. Until recently, functional MRI was implemented in AD and MCI diagnosis, which allowed voxel-wised whole brain study [23][24][25]. It helped clinicians to differentiate AD and MCI from HC. VD diagnosis still mainly relies on structural imaging, such as FLAIR and DW MRI. The application of fMRI in VD was mainly on priori based graph theoretical analysis [42], seed based analysis [43], or using particular band frequency [44]. The IFC analysis in VD was limited. Therefore, the aberrant VMHC maps presented in this study, together with the result in ROC analysis could be a potential biomarker to supplement the early diagnosis of VD.

VD vs. HC
Our result showed that IFC decreased in frontal orbital gyrus and gyrus rectus, but increased IFC in DMN including superior frontal gyrus, inferior frontal gyrus triangular, anterior and posterior cingulate gyrus; in precentral gyrus of ECN; in other parts of SN Previous studies suggested that vascular cognitive impairment encompasses all causes of cerebrovascular diseases, including silent brain infarct, leukoaraiosis, lacunar infarcts, medial temporal atrophy, and vascular comorbidities [12,45]. In this study, VD showed a significantly larger volume of leukoaraiosis, medial temporal atrophy, and higher vascular comorbidities when compared to the other three groups.
Leukoaraiosis has been reported as having a strong association with global cognition, executive function, and psychomotor response [46]. VD in this study demonstrated severe leukoaraiosis with significantly lowered HK-MoCA score compared to HC. Chen et al. (2019), demonstrated that severe leukoaraiosis associated positively with enhanced IFC within SN [47]. The anterior insula and anterior cingulate gyrus plays an important role in SN, which is modulated by homeostatic, cognitive, and emotional stimuli [48,49], and as a switch between ECN and the DMN across different stimulus modalities. Enhanced IFC in SN can be a compensatory action to maintain the regulation between ECN and DMN, which are vital to maintain daily activities. However, Chen's study used ICA to interpret functional connectivity, which required a priori region of interest for analysis. VMHC is a voxel-wise study of whole brain study of functional connectivity, all voxels were analyzed and provided comprehensive maps for comparison.
Additionally, our result demonstrated IFC decreased in frontal orbital gyrus, which encompass many different functions, including sensory integration to modulate behavior through the motor system, learning participation, and decision making for reward related behavior [50]. A previous study described that patients with dysfunction in frontal orbital cortex had impulsive responses, including deficits in motor response inhibition [51]. This may explain the increased IFC in the supplementary motor area to compensate for such deficits. The study of Liu et al. (2015) suggested that the IFC increased in bilateral motor cortex was to compensate the anatomical connection damage caused by the subcortical stroke [52]. Further investigation is needed to identify the correlation between the two regions.
Medial temporal atrophy represents volume loss in the hippocampal region [53]. It is one of the sensitive biomarkers identified for AD and associated cognitive impairment [54] and VD [55]. Our study result was similar to the Cho et al. (2009) study, which indicated the atrophy in AD was larger than that in VD. However, atrophy presents in later stage as the disease develops. For early diagnosis, upstream imaging techniques such as functional MRI are recommended as supplementary investigation.
In view of vascular comorbidities, the study of Schuff et al. (2009) suggested that the mentioned risk factors (T2DM, hypertension, and hyperlipidemia) were associated with cerebral perfusion decline [56]. Their study showed that both gray matter atrophy and cerebral blood flow (CBF) reduction coexist in frontal and parietal cortex in VD, similar to the location of IFC reduction in our study. It suggested that those risk factors might contribute to the IFC reduction in our cohort. Further investigation on CBF is suggested to investigate the association between IFC and CBF. In addition, Hachinski et al. (2008) suggested that cognitive impairment could result from severe microvascular brain disease [57]. The metabolic reduction showed in different anatomical regions between AD and VD, were similar to the regions of interest in our study, and might help to differentiate VD from AD.

AD vs. HC
Compared to HC, significantly lowered VMHC was observed throughout the limbic, frontal, parietal, temporal, occipital regions, that is DMN, ECN, and SN. Disruption of the network will have a negative effect on normal functioning and caused severe consequences. Our study showed similar results to the study of Wang et al. (2015), where they found that VMHC decreased in the orbital frontal cortex, anterior cingulate, putamen, caudate, insula nucleus accumbens, and primary olfactory cortex [24]. According to the findings in Braak and Braak 1997, Wang's study matched the definition of amyloid deposits during the intermediate stage of dementia (between stages A and B) [24,58]. The amyloid deposits were found in basal neocortex, including perihinal and ectorhinal regions initially. The depositions increase and spread to olfactory, anterior cingulate, orbitofrontal, caudate, insula, and hippocampal regions at stage B. Our result showed reduced VMHC in a more widely spread manner, in operculum, superior frontal gyrus, middle temporal gyrus, fusiform gyrus, inferior temporal gyrus, lingual gyrus, calcarine, cuneus, and occipital lobe at intermediate stage (stage B); also advanced stage (stage C) including posterior cingulate, precuneus, and paracentral gyrus (parietal lobe). It matched the definition of advanced stage of amyloid deposition (stage C). Results from both studies matched the description of amyloid deposition pattern described in the study of Braak and Braak (1997). It suggested that interhemispheric functional connectivity decreases with the accumulation of amyloid in AD.
In view of using VMHC value as a diagnostic test, Yang et al. (2018) used amplitude of low frequency fluctuations (ALFF) and fractional ALFF (fALFF) to differentiate AD from HC, They achieved an accuracy of 80.2% for AD vs. HC and 75.36% for MCI vs. HC using ALFF based ROC analysis; while using fALFF features, they achieved an accuracy of 89.1% and 70%, respectively, in ROC analysis [59]. Our results achieved accuracy that ranged from 72% to 81% using regional VMHC (81% for AD vs. HC, 72% for MCI vs. HC, and 78% for VD vs. HC). Our results were comparable to theirs even though we used a different method of analysis. In contrast to our study, they incorporated ALFF or fALFF values of 116 brain regions divided based on anatomical automatic labeling (AAL) template, and we combined regional VMHC where they were significantly different from HC. We both yielded higher accuracy when combining functional connectivity of different regions, indicating that functional connectivity of multiple regions improved the diagnostic test capability.

MCI vs. HC
MCI is a precursor to AD, and it has been classified as the prodromal stage of Alzheimer's disease [60]. In the current study, the IFC reduction was only confined to the anterior brain regions (gyrus rectus, olfactory, superior and inferior frontal orbital gyrus). Wang et al. (2015) showed a similar pattern of decreased VMHC in anterior brain regions including orbitofrontal, anterior cingulate, nucleus accumbens when compared to HC. However, increased VMHC was found in the sensorimotor cortex when compared to HC [24]. The anterior brain regions matched the description of early stage of amyloid deposition in the study of Braak and Braak (1997). While, the increases of VMHC in Rolandic operculum, superior frontal gyrus, insula, caudate, putamen, thalamus (amyloid spread to stage B) and precentral gyrus, paracentral gyrus, post central gyrus and supplementary motor area (amyloid spread stage C) follows the route of spread of amyloid deposition. The trend suggested that mild deposition of amyloid induces functional connectivity increases to compensate the loss of function [58]. Qureshi et al. (2019) used independent component analysis (ICA) to classify different levels of cognitive impairment and they achieved classification accuracy of 92.3% based on the rs-fMRI data. With the power of artificial intelligence, more rs-fMRI features can be included for analysis that may improve the diagnostic accuracy.
Other than using resting state functional MRI to study functional connectivity in detection of AD, MCI vs. HC, brain structural change has been used to differentiate AD and MCI from HC. Hippocampal atrophy was used and achieved AUC 74% and 65% for AD vs. HC and MCI vs. HC comparison, respectively, in Teipei et al. (2017) [61]. However, hippocampal atrophy usually exists in the late stage of AD. As the development from MCI to AD can last for years, it is not a good diagnostic tool for early AD or MCI detection. Furthermore, our results achieved higher AUC with balanced sensitivity and specificity. It indicated that VMHC is a better tool in differentiating AD and MCI from HC compared to using hippocampal volume reduction.

VMHC at Olfactory Network and Salient Network
Gyrus rectus lies medial to olfactory sulcus which located in the frontal lobe, and inferior to the olfactory frontal sulcus [62]. The olfactory network includes frontal orbital gyrus, olfactory cortex, and insula [63]. For all three diseased groups, the VMHC values in gyrus rectus and olfactory network were significantly lowered when compared to HC. A previous study showed no significant correlation between gyrus rectus and cognitive outcome in terms of MMSE, but subgroup analysis showed that language and memory recall were affected after surgical removal of gyrus rectus [64]. Lu et al. (2019) suggested that functional connectivity between the olfactory network to the right hippocampus were reduced in early stage MCI, late stage MCI and AD when compared to HC [65]. The level of impairment is directly related to the disease stage from early MCI to AD. Our result showed attenuated IFC in frontal orbital gyrus in VD but activated IFC in insula and olfactory. Insula is involved not only in the olfactory network, but also in the salient network. The function of the salient network (dorsal ACC and bilateral insula) appears to help in selection and segregation of relevant internal and external stimuli to guide behavior [66]. We hypothesized that the activated IFC in insula was related to promoting the function of the salient network to maintain daily functioning for VD patients by increased IFC in DMN and ECN.

Limitations and Future Study
Firstly, the sample sizes of this study in different diseased groups were small. Thus, our study result should be considered as a pilot study. Additionally, the number of participants with Lewy body dementia (DLB) and frontotemporal dementia (FTD) were too little for adequate evaluation. Secondly, due to the involvement of radiopharmaceutical injection, HC did not undergo F18-flutemetamol PET scan. Thirdly, vascular risk factors (e.g., diabetes status and lipid profile) of participants were obtained, but correlation analysis was not performed. Our current results were more exploratory and descriptive in nature. Additionally, detailed neuropsychological evaluation with impaired performance in different domains was not performed. It prevented further analysis on the correlation between regional VMHC and performance in neuropsychological domains in different diseased groups. As previously mentioned, accumulation of amyloid in brain regions may affect IFC and further study about the correlation of VMHC with amyloid accumulation can be conducted. This will form the pathophysiological basis of the diagnostic capabilities of VMHC to differentiate between AD, MCI, and HC. As discussed, cerebral blood flow, macro-and micro-vascular complications, and leukoaraiosis may be risk factors for VD. Combining VMHC and the above parameters may help to improve the accuracy to differentiate VD from other types of dementia. However, the duration and medication history of T2DM, hypertension, hyperlipidemia were not collected, which limited the analysis of severity level of abovementioned vascular comorbidities. Further studies performing correlation analysis of VMHC with CBF and the status of vascular comorbidities were suggested.

Conclusions
Our results showed that aberrant IFC was found in VD, AD, and MCI when compared to HC. VMHC maps showed specific patterns of reduction and increase of IFC in different dementia groups. Retrieving VMHC values in different brain regions can characterize AD, MCI, and VD from HC. Through combining the VMHC values at relevant brain regions, the diagnostic capabilities of AD, MCI, and VD from HC could be improved with accuracy of 92%, 83%, and 87%, respectively.