Elsevier

Neurobiology of Aging

Volume 29, Issue 1, January 2008, Pages 23-30
Neurobiology of Aging

Automated cortical thickness measurements from MRI can accurately separate Alzheimer's patients from normal elderly controls

https://doi.org/10.1016/j.neurobiolaging.2006.09.013Get rights and content

Abstract

We investigated the potential of fully automated measurements of cortical thickness to reproduce the clinical diagnosis in Alzheimer's Disease (AD) using 19 patients and 17 healthy controls. Thickness maps were analyzed using three different discriminant techniques to separate patients from controls. All analyses were performed using leave-one-out cross-validation to avoid overtraining of the discriminants. The results show regionally variant patterns of discrimination ability, with over 90% accuracy obtained in the medial temporal lobes and other limbic structures. Multivariate discriminant analysis produced 100% accuracy with six different combinations, all involving the parahippocampal gyrus. We therefore propose automated measurements of cortical thickness as a tool to improve the clinical diagnosis of probable AD, as well as a research method to gain unique insight into the etiology of cortical pathology in the disease.

Introduction

Brain imaging has played an increasingly important role in the study of dementias over the last decade. Within the realm of magnetic resonance imaging (MRI), work has predominantly focused on the segmentation of structures known to be involved in dementia, especially the parahippocampal gyrus and the hippocampus [21], [2], [20], [28], [5]. These methods have recently been complemented with techniques employing whole brain analyses [1], [18], [32], [6], [12], [24].

One of the key roles that imaging can play in dementia research as well as clinical practice is to aid the early diagnosis process (see [30], [5], [37] for reviews of the field). The crucial task is to identify image-based metrics that can accurately differentiate patients from normal elderly controls. Ultimately, the goal will be to discover a metric that can detect dementia at an earlier stage than a standard neurological diagnosis.

Discriminating normal elderly controls from patients diagnosed with probable AD using MR imaging techniques achieves accuracies ranging from 58% to 100% [5]. This wide range in accuracies reported indicates that the search for good discriminants continues. Moreover, almost all methods described to date involve manual measurements, which are time consuming, subject to inter-rater variability, and require expert anatomists in order to be performed correctly. Normal analysis is furthermore often restricted to a pre-defined subset of brain regions, which has a number of limitations. First is the theoretical challenge in performing differential diagnosis of the various dementias, since each have their unique anatomical presentations, featuring different rates and foci of cortical atrophy [14], [12]. The second issue is that using prior regions of interest restricts the use of additive discriminant models. For example, it is conceivable that the best discriminant is not just the volume of the entorhinal cortex, but rather a model containing the entorhinal cortex and prefrontal cortex volumes.

Discriminant analyses need to separate the training and validation datasets to be methodologically robust. Lack of such separation limits the generalizability to the population at large and risks classifier overtraining. Given an arbitrarily complex classification decision boundary any number of groups can be correctly separated; such boundaries are, however, unlikely to correctly classify new data not used in the creation of that boundary. Not separating training and validation datasets therefore results in inflated reports of accuracy. The training and validation datasets are ideally entirely separate; given a small study population, however, a leave-one-out or jack-knife approach can be employed. Here all but one of the members of the population is used to train the classifier, and the thus created decision boundary is used to classify the subject who was left out. The process is then repeated for all subjects, each being left out of the training once and subsequently classified. Such techniques avoid the danger of inflating accuracies while still retaining maximum power and should therefore always be used when performing discriminant analyses.

The objective of this study was to investigate the ability of automated cortical thickness measurements to discriminate Alzheimer's patients from normal elderly controls in the same subject population used in [24], using one time point per subject. Cortical thickness has ideal properties for this task: it covers the entire cortex, it is fully automated and thus not subject to inter-rater variability, and it reflects a fundamental neuro-anatomical property known to be affected in Alzheimer's disease [13], [15], [33]. Since there are theoretical grounds for using different discriminant techniques based on the task at hand (e.g. different covariance matrices between groups or different number of groups), three different tests were employed: (1) linear discriminant analyis (LDA), (2) quadratic discriminant analysis (QDA), and (3) logistic regression implemented through neural nets. In order to not over-train the classifiers, leave-one-out (jack-knife) cross-validation was performed throughout.

Section snippets

Participants

Thirty six subjects were studied. MRI scans were acquired from 19 patients (mean age 68.8 ± 6.9) and 17 controls (mean age 61.0 ± 9.1). The patients had the clinical diagnosis of probable AD according to the NINCDS-ADRDA [26]. Patients were recruited from the Department of Psychiatry, Alzheimer Memorial Center, Dementia and Imaging Research Group, Ludwig Maximilian University, Munich, Germany. Cognitive impairment in the AD patients was assessed using the Mini Mental State Examination (MMSE) [11].

Controlling for multiple comparisons

Multiple comparisons were controlled for using 1000 permutation tests. The fifth percentile of the resulting distribution – reflecting a p-value of 0.05 with strong control of Type I error – was found at an accuracy of 0.84%. No random permutation exceeded an accuracy of 0.85%.

Separating AD patients from normal elderly controls

Using mean thickness across the entire cortex to distinguish the patient and control groups resulted in an accuracy of 75%, sensitivity of 79%, and specificity of 71%. No single structure was able to perfectly separate

Discriminant ability of cortical thickness

The results show that cortical thickness was able to discriminate AD patients from normal elderly controls. The high accuracy obtained compares favorably with similar studies [6]. Strong control of multiple comparisons in this analysis suggests that any accuracies higher than 0.85 can strongly reject the null hypothesis of no accurate discrimination between patients and controls.

Previous studies have shown that accuracies range from 58% to 100% [5]. Apparent 100% discrimination inevitably

Conclusions

In summary, we presented a set of fully automatic techniques to extract cortical thickness from an AD cohort with normal elderly controls, and employed statistically well-controlled discriminant techniques at every point on the cortex to separate patients from controls. This allows for maps showing the distribution of accuracy, sensitivity, and specificity. Initial multivariate analyses combining automatically segmented cortical regions produced 100% accuracies in six different combinations,

Acknowledgements

ICBM grant PO1MHO52176-11, principal investigator Dr John Mazziotta; CIHR grant MOP-34996. Jason Lerch is funded by a K.M. Hunter/CIHR Doctoral Research Award.

Disclosures: The authors have no competing interests to disclose.

References (38)

  • D. MacDonald et al.

    Automated 3-D extraction of inner and outer surfaces of cerebral cortex from MRI

    Neuroimage

    (2000)
  • S. Robbins et al.

    Tuning and comparing spatial normalization methods

    Med Image Anal

    (2004)
  • P. Scheltens et al.

    Structural magnetic resonance imaging in the practical assessment of dementia: beyond exclusion

    Lancet Neurol

    (2002)
  • K.J. Worsley et al.

    Unified univariate and multivariate random field theory

    Neuroimage

    (2004)
  • H. Braak et al.

    Evolution of the neuropathology of Alzheimer's disease

    Acta Neurol Scand Suppl

    (1996)
  • G. Chetelat et al.

    Mapping gray matter loss with voxel-based morphometry in mild cognitive impairment

    Neuroreport

    (2002)
  • D.L. Collins et al.

    Automatic 3D model-based neuroanatomical segmentation

    Hum Brain Map

    (1995)
  • D.L. Collins et al.

    Automatic 3D intersubject registration of MR volumetric data in standardized Talairach space

    J Comput Assist Tomogr

    (1994)
  • A. Delacourte et al.

    The biochemical pathway of neurofibrillary degeneration in aging and Alzheimer's disease

    Neurology

    (1999)
  • Cited by (230)

    • An insight into applications of deep learning in neuroimaging

      2022, Artificial Intelligence for Neurological Disorders
    View all citing articles on Scopus
    View full text