Detection of Alzheimer's Disease using cortical diffusion tensor imaging

Abstract The aim of this research was to test a novel in‐vivo brain MRI analysis method that could be used in clinical cohorts to investigate cortical architecture changes in patients with Alzheimer's Disease (AD). Three cohorts of patients with probable AD and healthy volunteers were used to assess the results of the method. The first group was used as the “Discovery” cohort, the second as the “Test” cohort and the last “ATN” (Amyloid, Tau, Neurodegeneration) cohort was used to test the method in an ADNI 3 cohort, comparing to amyloid and Tau PET. The method can detect altered quality of cortical grey matter in AD patients, providing an additional tool to assess AD, distinguishing between these and healthy controls with an accuracy range between good and excellent. These new measurements could be used within the “ATN” framework as an index of cortical microstructure quality and a marker of Neurodegeneration. Further development may aid diagnosis, patient selection, and quantification of the “Neurodegeneration” component in response to therapies in clinical trials.

this "ATN" framework ("A" for amyloid deposition, "T" for tau levels, and "N" for neurodegeneration) AD forms a continuum in which the extreme points are represented by A−T−N−cognitively unimpaired subjects, and A+ T+ N+ subjects with dementia. The present study focused on change in the underlying neural architecture responsible for cognitive function as a potential "N" biomarker.
In addition to cell loss and synapse loss, the vertical cellular micro-circuits, known as minicolumns, which constitute the fundamental structure throughout the cerebral cortex, are altered in a graded manner during ageing, mild cognitive impairment (MCI), and AD (Chance et al., 2011). The microscopic disruption of columnar architecture correlates with plaque load and cognitive decline (van Veluw et al., 2012). A novel analysis method using Diffusion Tensor MRI to measure correlates of these cortical microstructural changes was previously validated against postmortem histology (McKavanagh et al., 2019) and tested in in-vivo cohorts (Dickstein et al., 2020;Torso, Ahmed, et al., 2020;Torso, Bozzali, et al., 2020).
The present study aimed to provide the first preliminary in-vivo validation of these neuroimaging measurements in AD cohorts to demonstrate that they are sensitive to dementia-related microstructural changes. This analysis method is complementary to other "N" biomarkers and requires only conventional MRI scanners, standard diffusion protocols, and no contrast agents. It is, therefore, potentially applicable to a variety of acquisition environments, including clinical.
This study aimed to test: (1) if the cortical diffusivity analysis provided generalizable in vivo measures of cortical grey matter diffusivity; (2) if the cortical diffusivity analysis can discriminate between groups; (3) how the discriminative power of the method compared with other clinical biomarkers (Cortical grey matter volume, AV45, and AV1451 PET for amyloid and tau).

| Study participants
A total of 78 individuals with probable AD and 71 healthy elderly controls (HC) from three different cohorts were included in the study.
The first cohort (24 AD, 23 HC) was an existing dataset recruited in Oxford (UK) (Zamboni et al., 2013) and was used as a "Discovery cohort" to explore the in-vivo validity of a novel method of cortical diffusivity analysis (Table 1).
The second cohort (29 AD, 23 HC) was an existing dataset recruited in Rome (Italy) (Giulietti et al., 2018) and was used as a "Test cohort" to test repeatability of the method in an independent sample (Table 1).
All subjects underwent extended clinical and neuropsychological assessments, which were centre specific (Giulietti et al., 2018;Zamboni et al., 2013), but included the Mini Mental State Examination (MMSE) (Folstein, Robins, & Helzer, 1983)  Rating scale (CDR) (Hughes, Berg, Danziger, Coben, & Martin, 1982 To control for the effect of head motion (Baum et al., 2018) in DTI maps, a displacement index generated using an in-house script was calculated (see Supporting Information). This value was used as a covariate in the General Linear Model (GLM) multivariate analysis.

| Cortical diffusivity analysis
The automatic cortical diffusivity analysis consisted of several differ- Readers may be familiar with MD as a measure of the total diffusion occurring in a voxel. It is calculated by finding an average of the three eigenvalues (i.e., [L1 + L2 + L3]/3). In the present study, additional measures were calculated as described in US20180143282A1: The perpendicular diffusivity was determined by multiplying the main eigenvector (V1) by the value of its corresponding eigenvalue (L1), then resolving this into its components. The value of the component perpendicular to the radial minicolumn direction across the cortex was the perpendicular diffusivity. Radial or parallel diffusivity was the component of the diffusion occurring in the principal diffusion direction that was parallel to the radial minicolumn direction across the cortex. The angle of columnar deviation, also called AngleR, was the difference between the radial minicolumn direction across the cortex, and the direction of the main eigenvector (V1), expressed as an angle.
The direction, CRadial, was derived, spanning the cortical ribbon between the pial and white matter boundary surfaces. Over 100,000 approximately evenly-spaced points on the white matter surface were taken, and cortical profiles were propagated through the cortical layers replicating the histological principles of radial minicolumns (Rakic, 1995), aiming to minimize the crossings of profiles, and reflecting the inside-out migration of cells along radial glial guidelines toward corresponding points representative of Cajal-Retzius cells at the pial surface. The cortical profiles were then selected for inclusion, taking into account features of cortical geometry that are known to influence or correlate with minicolumn width, shape and cell density, including cortical thickness and curvature.
All the cortical values were averaged to reduce the influence of noise in the DTI scans, effectively smoothing the data, and ensuring only directionality with some local coherence would dominate, therefore guarding against the influence of random deflections from the minicolumn direction. Each of the three novel metrics, AngleR, Per-pPD, and ParlPD, was based on an average from the whole cortex. As with other widely used metrics, such as whole brain volume (which does not discriminate between the many tissue compartments and sub-structures), a summary value for each subject has the advantage that it provides a good overview of group differences without the complications of sub-region sampling, requiring multiple covariates and multiple testing corrections.

| Validity
To test the validity of the method, the study design enabled the assessment of several different validity requirements: i. Repeatability: investigated as intrascanner variation, that is, the degree of variation produced by running the cortical diffusivity analysis on the same subjects (six controls) at two different time points, baseline and follow-up after a three-month interval, acquired on the same scanner.
ii. Reliability: investigated as Interscanner variation, that is the degree of variation produced by running the cortical diffusivity analysis on similar cohorts, acquired on different scanners. To do that, the cortical diffusivity analysis was run on the Discovery, Test and ADNI3 cohorts.
iii. Construct validity: the degree to which the cortical diffusivity analysis measured what it claimed to be measuring, was assessed using correlations between cortical diffusivity analysis measures and other common indices of brain structural degen-

| Diagnostic accuracy
To test the diagnostic accuracy of the cortical diffusivity analysis measures, different indices were estimated. The group discrimination capability (diagnostic group: HC vs. AD) of cortical diffusivity measures was investigated using Receiver Operating Characteristics (ROC) curve analysis and compared with a conventional diffusion measure (MD) and GM_fr (considered as a measure of atrophy and one of the main measures of neurodegeneration). As is well known, hippocampal atrophy is a sub-region value and is one of the main criteria to define AD diagnosis and therefore formed part of the group classification criteria, so it was not used in the discrimination capability comparison.
We considered as the "best discriminator" the feature with the  in the cohorts, using group membership as a fixed factor and head movement, subject age, and scanner as covariates. Differences between groups were tested with χ 2 -tests and t-tests. All the statistical results were thresholded at p < .05, after Bonferroni correction (0.05/number of comparisons).

| Statistical analysis
Pearson's and Spearman's correlations were used to investigate the associations among measurements. All p-values in correlation analysis were adjusted with false discovery rate correction (FDR <0.05) (Benjamini & Yekutieli, 2001).
3 | RESULTS In all cohorts, no significant difference was observed between groups for age, years of formal education, or sex.

| Demographics and clinical values
As expected, in all cohorts t-tests revealed lower MMSE (p < .0001) and higher CDR score (p < .0001) in the AD groups.

| Structural MRI analysis
Volumetric brain values are summarized in Table 2.
In the Discovery cohort, AD patients showed a significantly lower GM fraction than HC (t 45 = 4.457; p < .0001). As expected, the AD group showed a significantly lower Hipp Bil fr in (t 45 = 4.695; p < .0001). No significant between-group difference was found for the WMHs fr.
In the Test cohort, AD patients showed a significantly lower GM fraction (t 50 = 5.831; p < .0001). and Hipp Bil fr (t 50 = 6.934; p < .0001) than HC. Moreover, the t-test analysis showed a significantly higher WMHs fr in the AD group (t 50 = 5.253; p < .010).

| Cortical DTI analysis-results and validity
The results of the repeatability test, based on the comparisons between baseline and the 3-month follow-up, revealed that all cortical diffusivity measures (AngleR, MD, PerpPD, and ParlPD) had good to excellent ICC (α = .89-.93).
Concerning reliability (interscanner variation), the differences between HC and AD groups were tested in each cohort: In the Discovery cohort, the GLM showed that just the diagnosis had a significant overall effect (F 4,42 = 11.048; p < .0001  Table 3 for more details).

| Diagnostic accuracy (classification effectiveness in comparison to other methods)
The ability of each measure to correctly classify AD and non-AD subjects was assessed using the receiver operating characteristics (ROC) curve. Table 4 and Figure 2 show the principal measures of diagnostic reliability of cortical diffusivity measures. For each group, ROC analysis was performed on the structural volume measure (GM fr) and diffusion cortical indices (MD, AngleR, PerpPD, and ParlPD).
In the Discovery cohort, as expected, GM fr provided good group discrimination, having AUC = 0.816.
Therefore, additional statistics were generated for AngleR: Likelihood ratio values (LR− = 0.14; LR+ = 6.7) revealing that individuals with AngleR values greater than 0.981 θrad (Figure 3) had an increased probability of disease compared to individuals with lower AngleR values. Moreover, using the cut-off of 0.981 θrad all F I G U R E 2 Discriminations using the AngleR cutoff corresponding to the "best" point in the ROC curves F I G U R E 3 This figure shows that no single marker, including MMSE score, is adequate for identifying patients with a clinical diagnosis of Alzheimer's Disease. The x axis shows categories with increasing number of positive markers. In general, subjects with only one positive marker (toward the left side) are healthy controls (blue) and indicate false positives for the individual markers which report a positive. Whereas, with increasing combined marker positivity, more subjects are found to be AD patients (red) and on the right side of the graph the red points indicate false negatives for those markers which report a negative. Interestingly, a few AD patients were Amyloid negative subjects with above threshold MMSE, but were positive on the other markers. (Note that the ADNI3 protocol defined AD subjects with MMSE within the range 20-24, however, some subjects presented here were carried over from the earlier ADNI data sets within which the original criteria defined AD with MMSE 20-26) individuals were classified with an accuracy of 87% and a J of 0.74, NPV (87%) and PPV (88%).
In the Test cohort (similar to the Discovery cohort results), GM fr provided very good group discrimination having an AUC = 0.870.
AngleR obtained the best between-group discrimination, having an AUC = 0.931, followed by PerpPD (AUC = 0.903) and MD (AUC = 0.858). The AngleR cut-off point determined that the "best" point of the ROC curve was 0.983 θrad.

| CONCLUSION
Previous histological studies have revealed that AD results in progressive damage to minicolumn organization (Chance et al., 2011;van Veluw et al., 2012). This process, led by neurite loss and then neuronal death, causes progressive damage to the normal organization of cortical cells in columns, producing an alteration of cortical geometric properties (Chance et al., 2011;van Veluw et al., 2012). Therefore, we considered the alteration of such geometric properties as a biomarker, potentially measurable using a tailored, novel DTI analysis method.
Although DTI is a relatively crude tool for analyzing diffusion MRI, it can be useful for exploring markers of disease and has been shown to relate to the underling cytoarchitecture (McKavanagh et al., 2019).
With respect to the main aims of the study, the results suggested that the cortical diffusivity analysis did detect group differences accurately and satisfy validity requirements overall.
The validity of a test is based on its ability to measure reliably (Repeatability and Reliability) for the group of variables that it is designed to measure (Construct Validity) and to correctly distinguish subjects with the disease from healthy subjects (Diagnostic Accuracy) in accordance with other pre-existing scores (Concurrent validity). The validity of the method was tested here, to determine the possibility of generalizing the results obtained, by investigating repeatability, reliability, construct validity, and concurrent validity.
The intrascanner variation showed that the scores obtained at the two timepoints were strongly correlated and significantly consis- The structural MRI results are consistent with previous studies (e.g., Cuingnet et al., 2011), but the amyloid PET AUC differed slightly from some other studies (e.g., Palmqvist et al., 2015). Variation in PET amyloid results across studies could be due to sample size differences and/or selection of target regions. Both factors can produce significant changes in AUCs. The present study used the main whole brain amyloid value provided in the ADNI dataset.
There is, potentially, additive value in using a range of methods that provide complementary information and can provide increasing confidence of patient classification.
It is worth noting that these results are based on the diagnosis of moderate-severe AD, in order to explore the discriminatory power of the method on a well-characterized sample with clear diagnostic indices available. This enabled evaluation of the concurrent validity of the method. Of course, the ultimate goal is to move beyond the detection of moderate-severe AD, which offers limited insight for clinical practice, toward a preliminary validation of a method that could enhance quantification of the "N" component of the ATN framework (Jack Jr et al., 2018) earlier in the disease. This could have applications in differential diagnosis (Torso, Ahmed, et al., 2020;Torso, Bozzali, et al., 2020)) with respect to other forms of dementia and ideally in early diagnosis for detecting early changes in cortical architecture. As shown by previous studies (Dubois et al., 2016), the predictive power of conventional biomarkers in the preclinical AD population requires improvement, creating a need for new biomarkers and instruments capable of more effectively detecting preclinical AD.
The objective of the ATN criteria is to separate the biomarker profile of the disease that represents the underlying pathology from the clinical diagnosis of symptoms, which can often be mimicked in other forms of disease. This raises the prospect of a potential disconnect between the biomarker and the clinical definitions and evidence of this can be seen in Figure 3, where there are a number of HC individuals who may be in the preclinical stage of AD. It is also possible that "AD" individuals who are amyloid negative should be considered atypical. Nonetheless, the findings of the present study are broadly supportive of the principle of ATN criteria.

| LIMITATIONS
This study has some limitations: the study included a relatively small number in each cohort for the purpose of validation, further studies with larger cohorts would be recommended to fully generalize the findings. An additional limitation concerned interscanner reliability. In an ideal study the same individual subjects would be scanned using different acquisitions and on different scanners. This is very difficult to realize in practice, especially for subjects with AD, where a repeated acquisition would be very taxing and stressful for the patient, difficult to justify from an ethical point of view, and challenging for recruitment. All the data in the present study were drawn from existing datasets and in that respect at least, they do not represent a cohort specially optimized for our analysis.
In summary, the present study attempts to step towards building a bridge between previously characterized histopathological markers of dementia and current MRI methods. Further investigation on additional datasets will be needed, but this cortical DTI measurement, in addition to other methods already in use, appears to have the potential to contribute to improving diagnostic classification for Alzheimer's Disease. Such methods could form part of a repertoire of assessments to assist early diagnosis of the disease and differential diagnosis from other forms of dementia.