Jackknifing a normal database for regional cerebral blood flow SPECT obtained using principal components

A normal atlas for rCBF was created previously from 50 normal subjects using a mean image and several eigenimages to represent normal variation in rCBF SPECT images. The parameters of the atlas were optimized to differentiate between further normal subjects and patients with suspected dementia. These parameters included the optimal number of principal components. In this paper the Partial Correlations Test is used to determine the optimal number of principal components yielding similar results to that obtained experimentally. In order to determine the false positive rate, each image in the set is classified as normal or abnormal according to predefined criteria. This will be a biased classification as the image is included in the set used to define the normal atlas. The jackknife process is then applied to the normal database. This involves redefining the atlas by computing the mean image and eigenimages 50 times, missing out each individual in turn. The image of the individual is then reclassified on the basis of being excluded from the set. The results are similar to those obtained for images of a further 40 normal subjects obtained subsequently. The jackknife allows an unbiased estimation of the eigenimages which form the atlas. It is demonstrated that these vary very little from the originals, although there is some disagreement between the original and jackknifed eigenvalues.