Alzheimer’s disease (AD) is the most common form of dementia in the ageing population of today. The estimated cost of dementia worldwide has been calculated as 315.4 billion USD based on an estimated 29.3 million demented patients in 2005 (Wimo et al. 2007
). The number of patients with AD has been predicted to quadruple by 2050 (Brookmeyer et al. 2007
). The disease is characterized by a gradual loss of cognitive functions, such as episodic memory. The two major pathological hallmarks of AD are extracellular plaques and intracellular tangles. Plaques and tangles are built of aggregates of Aβ (Glenner and Wong 1984
; Masters et al. 1985
) and hyperphosphorylated tau (Goedert et al. 1991
), respectively. Other characteristics of AD are synaptic loss and neuronal cell death, leading to brain atrophy. Magnetic resonance imaging (MRI) provides structural information about the brain and has for many years been widely used for early detection and diagnosis of AD (O’Brien 2007
; Ries et al. 2008
). The way in which AD atrophy progresses through the brain has been described by Braak and Braak (1991
). Atrophy typically starts in the medial temporal and limbic areas, subsequently spreading to parietal association areas and finally to frontal and primary cortices. For many years studies have focused on single structures in the medial temporal lobe for the early diagnosis of AD, such as hippocampus and entorhinal cortex (Fox et al. 1996
; Jack et al. 1992
; Juottonen et al. 1999
). In recent years however, research has focused on combining different regions to look at patterns of atrophy instead of single measures and the former approach has proven to be more sensitive (McEvoy et al. 2011
; Westman et al. 2011c
; Zhang et al. 2011
). MRI is today an integrated part of the suggested research (Dubois et al. 2007
) and diagnostic criterion (McKhann et al. 2011
) alongside cerebrospinal fluid (CSF) markers and positron emission tomography (PET).
Freesurfer is a highly automated structural MRI image processing pipeline which produces regional volume, cortical thickness, gray matter volume, surface area, mean curvature, gaussian curvature, folding index and curvature index measures. Automated image analysis pipelines may have particular advantages when it comes to widespread uptake in either clinical or research practice. Manual measures of different brain regions are time consuming and operator dependent and therefore not always practical in a clinical settings. However, automated tools must be precise, accurate, fast and must be validated and tested on large cohorts. Several groups have utilized automated pipelines in AD research (Cui et al. 2011
; Li et al. 2011
; McEvoy et al. 2009
). We have also previously used automated image analysis pipeline output analyzed with multivariate tools for the purpose of AD classification and to predict conversion from the prodromal stage of the disease, mild cognitive impairment (Westman et al., 2011a
). Different regional MRI measures have been used in the studies reported in the literature including our own and different approaches have been taken to normalization. For example, should regional volumes be normalized by dividing by intracranial volume to reflect differences in head size between individuals, particularly males and females, and pre-morbid brain size? It is not clear yet which combination of regional measures and which normalization approaches yield the best results for individual classification and prediction.
The current study investigated the use of regional MRI measures analyzed by orthogonal partial least square to latent structure (OPLS) a multivariate tool for classification of individual subjects. The specific aims were to determine: (1) which type of normalization approach is most useful for the different regional measures (2) which combination of regional measures results in the best classification accuracy when distinguishing between AD subjects and healthy controls, and (3) to prospectively predict conversion from MCI to AD at baseline by appropriate choice of multivariate model. We hypothesized that regional volumetric measures would give the best results when normalized by total intracranial volume, that surface area should be normalized by whole brain surface area, while the remaining measures (cortical thickness, mean curvature, gaussian curvature, folding index and curvature index) should not be normalized. Further, we hypothesized that a combination of regional subcortical volumes normalized by intracranial volume and un-normalized cortical thickness measures would generate the most accurate predictions.