The results demonstrate that automated MRI measures of entorhinal cortex thickness, hippocampal volume and supramarginal gyrus thickness identify MCI and Alzheimer's disease individuals with excellent discrimination accuracy and specificity, exhibit a high degree of consistency and reproducibility across multiple independent cohorts, and correlate strongly with clinical measures of decline as well as cellular biomarkers. Taken together, these findings suggest the feasibility of using automated, MRI-based software tools as a diagnostic marker for Alzheimer's disease.
The regression analyses presented here indicate that automated MRI measures can differentiate MCI and Alzheimer's disease from normal ageing with excellent discrimination accuracy. In the comparisons between the MCI individuals and OCs, entorhinal cortex thickness, hippocampal volume and supramarginal gyrus thickness demonstrated an average AUC of 0.91 in the training cohort and an AUC of 0.95 in the validation cohort. Using these same MRI measures, patients with mild Alzheimer's disease could be differentiated from OCs with perfect discrimination (AUC = 1.0). These AUC values are more accurate than prior MRI (Xu et al
; Devanand et al
; Colliot et al
; Kloppel et al
), FDG–PET (Mosconi et al
; Jack et al
) or amyloid-binding PET studies (Jack et al
; Li et al
). The MCI discrimination accuracies presented here are comparable to one prior PET study utilizing a radioactive amyloid and tau protein tracer (Small et al
) and two prior MRI studies where a smaller number of subjects were examined from a single cohort (Killiany et al
; Davatzikos et al
). Further studies are needed to determine whether combining structural MRI measures with other imaging modalities will improve diagnostic and prediction accuracy and whether the benefits of using multiple methods outweigh the costs.
To the best of our knowledge, this is the first study to demonstrate that automated software tools can be utilized to quantify the atrophy of individual anatomic regions in a highly specific and precise fashion. The fact that entorhinal cortex thickness and hippocampal volume were two of the best discriminators of MCI indicates the specificity of these automated MRI methods for identifying the two regions implicated in the earliest stages of Alzheimer's pathology (Braak and Braak, 1991
; Kemper, 1994
). Consistent with prior MRI studies (Scahill et al
; Buckner et al
; Dickerson et al
), these results also highlight the relative importance of examining lateral parietal regions, such as the supramarginal gyrus, as important discriminators for the earliest stages of Alzheimer's disease.
The regression results further illustrate that these automated MRI measures are highly consistent and reproducible. In the comparison between the MCI individuals and OCs, both the training and the validation cohorts demonstrated similar AUC values indicating the reliability of these measures across multiple independent cohorts. Furthermore, for the validation cohort, the application of the training cohort logistic regression coefficients resulted in an AUC value of 0.95, the same as the value derived without the application of these coefficients. This shows that the model based on the training cohort using the three temporoparietal measures is clinically applicable and can be reproduced in populations other than that from which the training cohort were drawn.
The correlations between the three MRI measures and measures of clinical severity (i.e. CDR-SB and MMSE) suggest the potential for using these measures as surrogate markers of underlying disease. Correlations between tests of episodic memory function (AVLT 5 and 30 min recall) and measures of entorhinal cortex thickness and hippocampal volume are consistent with the fact that declines in episodic memory function are reported as predictors of disease progression. Future studies will examine whether combining these automated MRI measures with neuropsychological assessments will better predict which MCI individuals eventually progress to Alzheimer's disease.
Correlations between the three temporoparietal measures (i.e. entorhinal cortex thickness, hippocampal volume and supramarginal gyrus thickness) and CSF measures of tau, p-tau and abeta 42 suggest that these MRI measures are likely to be a reflection of known underlying Alzheimer's disease pathology. When considered together with the regression results, these data suggest the hierarchical fashion in which pathology affects the earliest stages of Alzheimer's disease, with tau-associated neurofibrillary changes in medial temporal regions and abeta-associated amyloid changes in the entorhinal cortex and neocortical regions (Arnold et al
; Braak and Braak, 1991
; Kemper, 1994
The methods we have described here can be implemented in clinical practice for the diagnosis of MCI and Alzheimer's disease. Using these software tools a single volumetric T1-weighted MRI scan can be completely processed, with little to no manual intervention, in a relatively short amount of time. The training cohort regression coefficients presented here can then be applied to the final output values of entorhinal cortex thickness, hippocampal volume and supramarginal gyrus thickness to calculate the predictive probability of a single individual being diagnosed as either MCI or Alzheimer's disease.
The present study has limitations. Since the MCI individuals in the two cohorts were diagnosed using slightly different criteria, differences between the two MCI groups could have affected the ability to independently assess the discrimination accuracy of the automated MRI measures. Another limitation is that the two MCI cohorts had differing percentages of males and females, with the training cohort comprised of a larger number of females and the validation cohort comprised of a larger number of males. The fact that the AUC values were comparable between the two cohorts and that the application of the training cohort logistic regression coefficients resulted in the same AUC value as without the application of these coefficients, suggests that the differences observed between the two cohorts did not play a major role in affecting the main findings of this study.
One concern is that although the procedures demonstrated here generalized across clinically diagnosed Alzheimer's disease and MCI populations, these procedures may be less accurate in the clinical setting where a range of cognitive disorders and dementia subtypes are present. The fact that the current results show complete discrimination suggests that these tools would be additionally powerful in the clinical setting. Future work will examine the application of these automated MRI measures to a larger, community-based, volunteer cohort that would be more representative of a clinical setting. Another concern is regarding clinical utility and whether these automated MRI measures can predict progression from MCI to Alzheimer's disease. Recently, we have completed a study examining the feasibility of using these automated MRI measures to identify those MCI individuals, within a larger MCI cohort, at greatest risk for Alzheimer's disease. Preliminary evidence from this study indicates that these automated MRI measures can identify MCI converters from MCI non-converters with a high degree of accuracy and have significant benefit when compared to clinical and neuropsychological assessments alone for predicting progression from MCI to Alzheimer's disease (Desikan et al
The identification of individuals in a transitional phase is critical for testing disease-modifying therapies and for the development of novel medications to prevent or delay Alzheimer's disease. The results from this study demonstrate that automated MRI-based neuroanatomic measures provide one cost-effective and efficient method to identify individuals in the earliest stages of Alzheimer's disease and may further serve as a quantitative and biologically meaningful endpoint in therapeutic trials.