The purpose of this research was to improve the accuracy of magnetic resonance imaging for diagnosing Alzheimer's disease. Prior to building an image-based multivariate pattern classifier, the authors transformed the image-derived features so that they reflected variability in a healthy control population. This step improved the classifier's diagnostic accuracy.

The most common feature-standardization approaches utilize data from both healthy and diseased groups to transform each image-derived feature, before using them to build a diagnostic-pattern classifier. Using this total variation weakens the separability of the data in high-dimensional space, and as a result, differences between the groups are more difficult to detect. In this work, the authors proposed an alternate approach that uses an estimate of the healthy control-group standard deviation to transform the image-derived features before training the classifier. Their method led to improvements in classifier performance—and hence more accurate diagnoses of Alzheimer's disease from magnetic resonance images of the brain.

Read the article in Neuroimage.