Individual patient diagnosis of AD and FTD via high-dimensional pattern classification of MRI

Abstract

The purpose of this study is to determine the diagnostic accuracy of MRI-based high-dimensional pattern classification in differentiating between patients with Alzheimer's disease (AD), Frontotemporal Dementia (FTD), and healthy controls, on an individual patient basis. MRI scans of 37 patients with AD and 37 age-matched cognitively normal elderly individuals, as well as 12 patients with FTD and 12 age-matched cognitively normal elderly individuals, were analyzed using voxel-based analysis and high-dimensional pattern classification. Diagnostic sensitivity and specificity of spatial patterns of regional brain atrophy found to be characteristic of AD and FTD were determined via cross-validation and via split-sample methods. Complex spatial patterns of relatively reduced brain volumes were identified, including temporal, orbitofrontal, parietal and cingulate regions, which were predominantly characteristic of either AD or FTD. These patterns provided 100% diagnostic accuracy, when used to separate AD or FTD from healthy controls. The ability to correctly distinguish AD from FTD averaged 84.3%. All estimates of diagnostic accuracy were determined via cross-validation. In conclusion, AD- and FTD-specific patterns of brain atrophy can be detected with high accuracy using high-dimensional pattern classification of MRI scans obtained in a typical clinical setting.

Fig. 1

Representative cross-sections of effect sizes calculated for CN minus FTD (top), CN minus AD (bottom). No negative effect sizes were measured for either of the 3 comparisons. Effect sizes above 6 were all mapped to the same color (dark red), and effect sizes less than 3 are not shown. Images are in Radiology convention.

Fig. 2

Effect sizes of the difference FTD-AD GM RAVENS maps. The two rows show two different sections through the brain highlighting fronto-temporal (top) and parietal (bottom) regions of reduced volumes. The two left columns are from all AD patients, whereas the right two columns are from relatively younger AD patients that were age-matched to the FTD patients. Images are in Radiology convention.

Fig. 3

Representative cross-sections of effect size maps showing regional volume differences between CN and FTD patients in regions displaying some inter-hemispheric asymmetry (arrows). Top: white matter. Bottom: gray matter. Some degree of left>right atrophy in the temporal lobe, and right>left atrophy in some frontal lobe areas is evident, albeit it is relatively small. Images in Radiology convention.

Fig. 4

ROC curve showing the sensitivity and specificity in differentiating between FTD and AD. Sensitivity here reflects the percent of FTDs that were correctly classified as FTD against a background group of AD, and specificity reflects the percent of AD participants correctly classified as AD. The ROC curve was obtained by varying the classification threshold applied to the SVM decision function.

Fig 5

Scatter-plots of volumes obtained from local clusters placed on regions of highest effect size between FTD and AD, identified via voxel-based analysis of the tissue density maps of GM: left and right orbitofrontal cortex, and right entorhinal cortex (ERC).

Fig. 6

Scatterplot of the hippocampal volume against the total brain volume (without the cerebellum and brain stem), for FTD, AD, and their respective cognitively normal controls. These measures are commonly used in AD. Although group differences between both demented groups and their controls are measured, especially for the AD group, individual classification in insufficient using these measures, due to the overlap of the scattergrams.