IC‐P‐016: Regression‐to‐pons normalization of FDG‐PET improves discrimination of Alzheimer's disease from healthy aging
Year: 2010
Type: article
Abstract: Regional normalization of cerebral metabolism rates of glucose (CMRglc) is an important preprocessing step used to standardize fluorodeoxyglucose positron emission tomography (FDG-PET) images. In statistical comparisons of Alzheimer's disease (AD) and healthy aging, this normalization is typically done relative to the cerebral global mean or to the mean of an unaffected region, such as the pons. Global normalization is problematic because overall cerebral metabolism is abnormally low in AD. Currently, normalization to pons is done by voxelwise division by the average metabolism in the pons. This procedure is also problematic because the average pons value is inherently noisy, and this can increase the variance of the normalized voxel values. We propose to normalize FDG-PET images using linear regression of the CMRglc at each voxel versus the average pons value and to test if this normalization improves classification between AD and healthy aging. Unlike division, linear regression decreases the noise variance while accounting for the variability in pons metabolism. We evaluated three normalization schemes, global cerebral, division-by-pons, and the proposed regression-to-pons, in a support vector machine (SVM) classification of normal controls versus AD patients from the Alzheimer's Disease Neuroimaging Initiative (ADNI). We trained a classifier for each normalization scheme on 60 control subjects and 50 AD patients and then tested the classification on a separate test set including 21 control subjects and 21 AD patients. All analysis was done on 3D stereotactic surface projection (SSP) maps of glucose metabolism generated by Neurostat, which also computes the average pons values used. We used the global cerebral normalization that is part of the standard ADNI preprocessing. Classification using global normalization resulted in 81% specificity, 76% sensitivity, and an overall classification accuracy of 78.57%. Results for normalization by pons division were similar: 71% specificity, 81% sensitivity, and 76.19% accuracy. The proposed normalization by regression-to-pons achieved the best classification results: 95% specificity, 81% sensitivity, and 88.10% accuracy. Classification between AD patients and healthy controls is improved in specificity and overall accuracy when FDG-PET values are normalized using regression-to-pons. The proposed normalization scheme should have wide applicability, as variance reduction is a desirable property in many statistical analyses.
Source: Alzheimer s & Dementia
Language: English
Subfield: Psychiatry and Mental health
Field: Medicine
Domain: Health Sciences
Open Access status: bronze
OpenAlex ID: W2039358158