Radial Diffusivity is the Best Global Biomarker Able to Discriminate Healthy Elders, Mild Cognitive Impairment, and Alzheimer's Disease: A Diagnostic Study of DTI-Derived Data
Correspondence Address: Source of Support: None, Conflict of Interest: None DOI: 10.4103/0028-3886.284376
Source of Support: None, Conflict of Interest: None
Keywords: Aging, Alzheimer′s disease, diffusion tensor imaging, discriminant analysis, mild cognitive impairment
During the past two decades, several diffusion tensor imaging (DTI)-derived biomarkers have been assessed for the diagnosis of Alzheimer's disease (AzD). These parameters supplemented the characterization of the microstructural integrity of gray matter (GM) and white matter (WM) between healthy elderly brains and AzD, and also helped to diagnose a transitional state named mild cognitive impairment (MCI)., Since the last 15 years, open-source software can perform the quantitative analysis of global and regional microstructural WM and GM changes using DTI.
Although recent studies recognize up to 11 tensor metrics derived from DTI, each one with different diagnostic performance;, a common finding in most of the studies published on this topic is a limited measurement of DTI-derived metrics (the most commonly reported are fractional anisotropy, mean diffusivity, and axial and radial diffusivities).
For explanatory purposes, DTI-derived metrics can be classified into three groups: anisotropies; this group include fractional (FA) and relative (RA). Diffusivities, they consider simple isotropic diffusion (p), simple anisotropic diffusion (q), mean diffusivity (MD), radial diffusivity (RD), and axial diffusivity (AD). The tensors group include the total magnitude of the diffusion tensor (L); and the linear (Cl), planar (Cp), and spherical tensors (Cs).,
These biomarkers are calculated from DTI, which is an imaging sequence acquired using magnetic resonance imaging (MRI). DTI biomarkers can be used not only in a regional but also in a global (whole-brain) approach to differentiate healthy from diseased brains., A comprehensive approach to brain imaging is particularly helpful to understand the complex relationships among normal ageing, MCI, and AzD. For instance, it is still not clear if WM can be directly affected in AzD independently of whether WM changes mainly represent secondary effects of GM atrophy or GM degeneration; investigators agree that more research and validation studies are needed before it is realistic to use DTI information in clinical practice with individual patients.
Therefore, considering all the studies mentioned above, we conducted a feasibility study to assess if, recently DTI parameters, would be able to differentiate elderly healthy brains, MCI, and AzD, using a discriminant multivariate (DA) model.
Subjects and demographics
A retrospective study was performed using the MRI database of patients previously recruited from the Geriatric Unit of the first author of this study, from July 2011 to December 2012. All participants were right-handed subjects and classified as healthy elders, MCI, and AzD patients. They received complete physical and geriatric examinations by a board-certified geriatrician and neuropsychological evaluation as previously described. Clinical diagnosis followed the guidelines from the National Institute on Aging-Alzheimer's Association workgroups., Exclusion criteria included the history of significant neurological, psychiatric, or severe cardiovascular diseases. The Institutional Review Board approved the study and patients received an informed consent before underwent brain MRI. The study included 34 subjects: 27 females and 7 males. Twelve subjects were healthy elders volunteers: mean age, 70.4 ± 3.9 years (range, 66–80 years). Eleven subjects were detected with MCI (mean age, 74.4 ± 7.1 years; range, 63–84 years) and 11 patients had a diagnosis of AzD (mean age, 74.8 ± 11.461 years; range, 53–94 years).
Brain image acquisition
Brain MRI studies were acquired using a 3.0T Signa HDxt scanner (GE Healthcare, Waukesha, WI) and a high-resolution eight-channel head coil (Invivo, Gainesville, FL). Contraindications of the MRI study included the presence of a pacemaker or metallic implant and claustrophobia.
Conventional standard clinical sequences included the following: sagittal T1-weighted fluid-attenuated inversion recovery (FLAIR) (TE/TR = 10/2,500 ms) with a 5–3 mm slice thickness/gap, an FOV of 24 × 24; an axial Spoiled Gradient Echo (FSPGR) (TE/TR = 3.9/9.4 ms) with a 1.3/0 mm slice thickness/gap, an FOV of 24 × 18; a coronal T2-weighted fast spin-echo (FSE) (TE/TR = 164.1/2617 ms) with a 3/0 mm slice thickness/gap, an FOV of 22 × 16; and an axial FLAIR (TE/TR = 115.8/11002 ms) with a 5/1 mm slice thickness/gap, an FOV of 22 × 22. Axial DTI acquired 30 slices covering the entire brain and brainstem with 1.87 × 1.87 × 5.0 mm3 voxel size with 25 noncollinear directions. We used a b-value of 1,000 s/mm2, and another b-value of 0 s/mm2. We excluded structural abnormalities, such as a tumours or stroke, anatomical variations (e.g., mega cisterna magna, cavum septum pellucidum), or technical artifacts. WM hyperintensities on FLAIR and T2-weighted images were rated using the age-related WM changes (ARWMC) score, we excluded elderly subjects with a regional ARWMC score higher than one. DICOM images were managed using RADSpa version 3.5.2 (Telerad Tech Pvt. Ltd., Minnetonka, MN, USA). It is a web-based RIS, and PACS platform that is FDA approved and HIPAA compliant.
DTI analysis and global measurements
To convert Dicom images and to extract the quantitative information we used Dcm2nii, FMRIB FSL Library v. 4.1.9, and Tract-Based Spatial Statistics (TBBS) v1.1. Images from DTI were obtain using BET v. 2.1. Eddy currents were corrected using FMRIB's Diffusion Toolbox v. 2.0;Reconstruct Diffusion Tensor (DTIFIT), and each mean value of global tensor metric wascalculated from the eigenvector and eigenvalue maps using fslmaths tool; detailed methodology has been described previously. [Figure 1] shows the eigenvalue maps of nine out of the eleven DTI tensor metrics calculated for each patient.
Sample size and descriptive statistics
We included at least ten subjects per group and used a minimum overall sample size of 30. To summarize the data, we presented means ± SD; 95% and calculated confidence intervals (CI), we used the bootstrapping method with bias-corrected and accelerated confidence estimate that considered 1,000 bootstrap resamples. We used the Kruskal–Wallis test to test for differences among groups.
DA is optimal under the same conditions where Manova is optimal; we did not find deviations from Manova assumptions that might distort the statistical significance. Box's M-test assessed the homogeneity of variance–covariance matrices. Because of the similarity among tensor-metric formulae, we ran scatterplots and correlations to check the strength of relationships in the dependent variables to detect multicollinearity [Table 1]. Partial correlation analyses calculated the Pearson's correlation coefficient (r) and controlled the effect of age, gender, and clinical diagnosis. The association between correlation coefficients was graded as very strong (≥0.8), moderately strong (0.6–0.8), fair (0.3–0.6), and weak (<0.3). A squared r-value represented the coefficient of determination, the proportion of variance that each two compared variables had in common.
We performed the DA using the stepwise method that considered the value of Wilk's lambda and changing criteria: minimum partial F to entering of 3.84 and minimum partial F to remove of 2.71. Continuous variables identify specific tensor-metric attributes in the brains of subjects. The dependent variable (DV) was the categorical variable “clinical diagnosis” with three levels: healthy elders, MCI, or AzD. We included 12 independent variables (IVs) in our model: the DTI parameters summed 11 variables, and the 12th variable was the patients' age. The effect size was calculated using the squared canonical correlation as the equivalent of the R2 in regression with the calculation of their corresponding effect sizes. The model diagnostic evaluation was tested using a cross-validated contingency table generated by the DA model.
The area under the curve (AUC) was considered the measure of the overall performance. We perform calculations of sensitivity, specificity, predictive values, and likelihood ratios. The accuracy was graded as 0.90–1.0 excellent (A), 0.80–0.90 0 good (B), 0.70–0.80 0 fair (C), 0.60– 0.70 poor (D), and 0.50–0.60 fail (F).
Software. We used the IBM® SPSS® Statistics software (version 126.96.36.199 IBM Corporation; Armonk, NY, USA) and MedCalc bvba version 14.8.1 (Mariakerke, Belgium). The study followed the STARD initiative. A P < 0.05 indicated statistical significance.
Tensor-metrics comparisons across groups
We found a statistically significant difference in 6 out of the 11 tensor-metrics in the three clinical groups: MD, p, L, RD, AD, and Cp. For the first five significant variables, the AzD group recorded a higher mean rank score than the other two groups; for the Cp metric, the AzD group recorded the lower mean rank score; in all cases: df = 2, n = 32, and P < 0.05. [Table 1] presents the mean values, SD, and bootstrap CI of the DTI metrics in each group.
Partial correlation analyses
Among the pairs of bivariate correlations, we found 15 with a significantly very strong R-value (>0.8): MD ⇔ L (+), MD ⇔ RD (+), MD ⇔ AD (+), p⇔ L (+), p⇔ RD (+), p⇔ AD (+), L ⇔ RD (+), L ⇔ AD (+), Cl ⇔ Cp (+), Cl ⇔ Cs (+), Cl ⇔ RA (+), Cp ⇔ Cs (−), Cp ⇔ RA (+), Cs ⇔ RA (−), and RD ⇔ AD (+). [Figure 2]a depicts the correlation matrix and scatterplot of the 11 tensor-metrics.
Although some bivariate correlates showed R-values >0.8, we included those variables in the DA because we used the stepwise variant of DA that defends against multicollinearity and singularity. Box's M-value = 1.756, F = 0.836, df (1559.484), P = 0.434, this results indicated a nonsignificant assumption of homogeneity of variance–covariance matrices. In the stepwise statistics, RD was the only variable entered at Step 1: F = 9.746 (2, 27) P = 0.001; the other variables were not selected by the software to participate in the analysis.
Because of the stepwise modality of DA, only the first canonical discriminant functions were used in the analysis; it significantly differentiated across clinical groups: Wilks' λ = 0.581, χ2 (2) = 14.673, P = 0.001. Wilks' lambda delivered a reasonable proportion of total variability not explained by the model of 58.1%, and then a canonical correlation of 0.648 suggested that the model describes a 41.99% of the variation.
The Tests of equality of group means provided statistical evidence of significant differences among means of healthy elderly brains, MCI, and AD groups for only five of the IVs (MD, p, L, RD, and AD) with RD producing the highest F's value; [Table 2].
Standardized Canonical Discriminant Function Coefficients showed the importance of each predictor, a positive or negative sign indicated the direction of the relationship. A significant increase in RD was the only diagnostic predictor: AzD > MCI > normal elders. The coefficient of the selected variable stood out (for these data) as those that actively predict the allocation to each clinical group. For this coefficient score, the rest of the variables showed decreasing values as diagnostic predictors; [Table 3]A.
The structure matrix correlation depicted the relative rank of the predictors. This matrix correlation is considered more accurate than the Standardized Canonical Discriminant Function Coefficients. The largest loadings for each discriminate function suggest diffusivity patterns as the features that discriminate between groups. A cut-off value of 0.30 separates between relevant and less important variables. The RA and the tensors (Cp, Cs, Cl) had values below 0.30; [Table 3]B shows the unstandardized coefficients of the Canonical discriminant function coefficients table; these coefficients can be used to create a regression equation. The equation of the regression model in this study was:
D = [−11.473 × (Constant)] + [9,544.153 × (Radial Diffusivity)].
[Table 3]C indicates the partial contribution of each variable to the discriminate function controlling for all other variables in the equation.
The values of group centroids used the group means of the predictor variables called centroids: healthy elder = −0.773, MCI = −0.132, and AzD = 1.163. The cut-off value is considered the average between every two centroids: cut-off between healthy elder and MCI = −0.452 and cut-off between MCI and AzD = 0.515.
We finished the DA performing a classification phase. In the initial assessment, 74.2% of original grouped cases were correctly classified, after the cross-validated set of data only 64.5% of patients were accurately classified; this value corresponded to the overall predictive accuracy of the discriminant function. The mean of D scores for each group and the centroids for each cluster visually demonstrated the effectiveness of the discriminant function. Scatterplots and box plots of the average D scores for each cluster visually showed the efficiency of the discriminant function, the absence of overlap of the plots revealed excellent discrimination, [Figure 2]b and [Figure 2]c.
Diagnostic performance of the discriminant model
The AUC, the overall performance for each diagnosis, was competent (0.81) in AzD; fair (.73) for healthy brains; and weak (0.66) for MCI; the CI values evinced significance. [Table 4] presents a summary of the diagnostic tests.
AzD is the most common cause of dementia and was initially described by Alois Alzheimer in 1906. Other common dementing illnesses include vascular dementia, dementia with Lewy bodies, and frontotemporal lobar degeneration. Most cases of AzD appear after a period of MCI. The National Institute on Aging-Alzheimer's Association recently incorporated MRI and cerebrospinal fluid (CSF) as part of five accepted biomarkers for diagnosis. These biomarkers depict an increase of amyloid β accumulation (using PET or CSF) or neuronal injury (elevated levels of CSF tau, decreased FDG uptake, and brain atrophy on structural MRI). Several reports in the literature prove the usefulness of DTI metrics as diagnostic biomarkers.
Our findings acknowledge the feasibility of using only one global DTI metric, the RD. Our model can discriminate AzD from healthy elders and MCI with an AUC up to 81%. The way a clinician could use our predictive model is: when the mean discriminant score of a new patient is less than or equal to the lower cut-off, the case is classified as 1 (healthy elderly brain). When the discriminant score is above the first centroid but below the second centroid, it scores as 2 (MCI brain); and if it is above the value of the second centroid, it is classified as 3 (AzD brain). The merit of our model is the finding that only one parameter, RD, might be able to discriminate among clinical groups. However, the overall accuracy of our model, just 64.5%. As a feasibility study, we consider this report a first step in supporting the clinical use of DTI parameters to diagnose degenerative brain diseases using quantitative MRI. We acknowledge this model is far from perfection, as it only explains 41.99% of the variation in the grouping variable.
In defence of our global approach instead of focal measurements of atrophy in specific gyri, several studies revealed GM, and WM affection with widespread DTI changes in MCI and AzD compared with controls., We do not know why RD was the only elected metric in our study. RD represents the diffusivity perpendicular to the principal eigenvector (λ1, axial diffusivity); it is more likely affected by myelination, axonal diameter, and axonal packing. However, several studies have reported similar findings in line with our results: increased RD in temporal, frontal, and parietal regions in AzD. The uncinate fasciculus in AzD patients has shown an increase in RD over a period of 1.5 years compared with baseline. Widespread cross-sectional DTI differences between MCI and controls in RD have been detected at baseline and follow-up after 2.5 years. Our finding of increased global RD is similar to a regional increase of RD associated with CSF biomarkers that study found that RD increases and FA decrease in posterior cingulate for MCI patients with high pathological CSF total Tau levels when they were compared with MCI patients with nonpathological CSF total Tau levels. Similarly, patients with pathological levels of CSF Tau had greater FA reductions and RD increases in the right cingulum and superior longitudinal fasciculus compared with controls over time, but this finding was not detected in MCI patients with nonpathological CSF Tau levels.
We acknowledge several limitations: in the current knowledge about DTI, it is unclear how the tensor metrics values evolve. Because there are no studies that evaluated global DTI metrics,,, it was not possible to compare our results with others in the literature. The prospective use of RD as a single imaging biomarker is a delicate topic; it would mean that an individual DTI value of the whole brain (encompassing brain heterogeneities) might support a clinical diagnosis. Recent studies have proposed the use of diffusion kurtosis imaging (DKI) (an extension of DTI) as a more accurate method given the probable non-Gaussian diffusivity in human tissue; however, we did not include DKI in our analysis as it is a method that is still waiting for validation. About our statistical analysis, in situ ations where some DTI-metric values show high correlations (R >0.8), those variables provide redundant information making a matrix inversion unreliable; then, the usual solution is the deletion of the additional variable. However, if a researcher wants to retain all variables in her analysis, the IBM® SPSS® Statistics software protects against singularity and multicollinearity by using computation of pooled within-cell tolerance for each variable; this procedure is part of the stepwise method in DA.
Because the amyloid accumulation may precede the onset of clinical symptoms by 15 or more years, researchers need the patience to collect clinical follow-up data over an extended period, before making reliable conclusions on the biology of AzD. When the aim is to identify the earliest markers of the disease, the conversion may take several years, at least ten years in some cases, making such studies rare. Researchers need to be cautious when interpreting simple DTI metrics; there is still no consensus on how to attribute underlying biological processes to changes in the measured diffusion. Although FA and MD have been the most often employed measures of WM degeneration in early AzD, evidence suggests that different DTI metrics (AD and RD mainly) can be independently affected during the axon disruption.
The initial evaluation of patients with suspicion of dementia should include a conventional brain MRI before DTI's metrics calculation; several factors can be responsible for cerebral volume loss and WM and GM changes. Among these factors, we can mention the presence of surgically treatable lesions such as a mass lesion, subdural hematoma, or hydrocephalus; small and large vessel disease; cerebral microbleeds; patterns of general cortical atrophy; and focal atrophy affecting the frontal and temporal lobes, precuneus, hippocampi, midbrain, and pons.
In conclusion, global RD detects alterations in the GM and WM microstructure that can be used to supplement the classification of subjects as healthy elderly, MCI, or AzD. DTI can strengthen conventional brain MRI in the same evaluation, with a little extra cost when compared with nuclear medicine imaging or cerebrospinal fluid analysis. However, there is an urgent need for consensus about DTI acquisition parameters in regional and global approaches; and also in the timing to schedule MRI follow-ups. DTI biomarkers might have the potential to monitor disease progression, measure outcome in drug trials, and screen pre-AzD subjects. Additional studies and validation of data are needed before it is realistic to use this information with individual patients on a day-to-day basis.
Ethical standards and patient consent
We declare that all human and animal studies have been approved by the Institutional Review Board and have, therefore, fulfilled the ethical standards laid down in the 1964 Declaration of Helsinki and its later amendments. We declare that all patients gave informed consent before inclusion in this study.
Medica Sur Clinic and Foundation partially supported this study.
Haydee Garcia-Lazaro was scholarship recipient of the Instituto de Ciencia y Tecnologia del Distrito Federal of Mexico (awardee number EJE11-86) and worked as a research fellow at the MRI unit of Medica Sur Clinic and Foundation from 2011 to 2012.
David Cortez-Conradis worked as a research fellow at the MRI unit of Medica Sur Clinic and Foundation from 2012 to 2014.
Manuel Martinez-Lopez was Chairman of the MRI Unit of Medica Sur Clinic and Foundation from 1996 to October 2015.
Ernesto Roldan-Valadez was Coordinator of Research at the MRI Unit of Medica Sur Clinic and Foundation from 2010 to April 2015.
Financial support and sponsorship
Conflicts of interest
There are no conflicts of interest.
[Figure 1], [Figure 2]
[Table 1], [Table 2], [Table 3], [Table 4]