Computing Publications

Publications Home » Multi-region analysis of longitud...

Multi-region analysis of longitudinal FDG-PET enables accurate AD classification

Katherine Gray, R. A. Heckemann, A. Hammers, Daniel Rueckert

ICAD 2011
July, 2011

Background: Imaging biomarkers for AD are desirable for improved diagnosis and monitoring, as well as drug discovery. The Alzheimer's Disease Neuroimaging Initiative (ADNI) is a valuable resource for related investigations, providing longitudinal clinical and imaging data from patients with AD and mild cognitive impairment (MCI), as well as healthy controls (HC). ADNI data have been used to demonstrate the value of incorporating longitudinal information from MRI, but there has been no corresponding classification study involving longitudinal FDG-PET. We therefore investigate the value of combining baseline and month-12 FDG-PET information for classification.

Methods: Image data are from 221 ADNI participants for whom baseline and month-12 images were available: 50 AD, 54 HC and 117 MCI, of whom 53 have progressed to AD (pMCI), whilst 64 so far remain stable (sMCI). Baseline and month-12 MRIs were automatically segmented into 83 anatomical regions. Regional signal intensities were extracted from baseline and month-12 FDG-PET images, and regional percentage changes in signal intensity over 12 months were calculated. Global variations in the cerebral metabolic rate of glucose between subjects were accounted for using a cluster covering areas of the brain that are relatively unaffected by AD. The regional features were provided to a support vector machine classifier, and robust estimates of classifier performance were assessed using bootstrap resampling.

Results: Highly significant increases in classification accuracy are achieved when using month-12 signal intensities compared to using baseline signal intensities. Accuracies increased from 81% to 86% for AD/HC, 71% to 79% for pMCI/HC, 73% to 79% for AD/sMCI, and 58% to 62% for pMCI/sMCI. More interestingly, further significant increases in accuracy may be achieved by combining month-12 signal intensities with the percentage changes over 12 months. Using this feature combination yields accuracies of 88% for AD/HC, 81% for pMCI/HC, 84% for AD/sMCI, and 63% for pMCI/sMCI.

Conclusions: These results surpass many state-of-the-art image-based classification methods. This study demonstrates that information extracted from serial FDG-PET through regional analysis can accurately discriminate diagnostic groups, a finding that may be usefully applied in the diagnosis of AD, predicting disease course in individuals with MCI, and in the selection of participants for clinical trials.


Poster available at

BibTEX file for the publication built & maintained by Ashok Argent-Katwala.