Optimization of factors affecting the state of normality of a medical image

The purpose of this paper is to examine the first stage of the diagnostic process in medical imaging, namely determination of the state of normality, and to attempt to optimize factors contributing to this stage. An image of a given type is defined as abnormal if it does not belong to the appropriate class of normal images. All images must be pre-processed involving image registration and normalization to align and scale the images with respect to each other. Normal ranges may be determined for each voxel (or other appropriate region) from a representative normal sample using univariate analysis, obtaining mean and standard deviation images, or multivariate analysis, which accounts also for normal patterns of variation (represented as principal components). For a new image, the variation from normality (in SDs) for each region may be determined. Since the spatial distribution of this parameter is thought to be relevant, connectivity of abnormal voxels was considered as a possible factor. For the purposes of this study, SPECT images indicating regional cerebral blood flow were used. Images from 50 normal subjects formed the normal sample. A further 40 normal subjects and 200 patients referred with suspected dementia were then analyzed using the normal ranges. ROC analysis, using number of SDs as a variable threshold, was used to optimize the factors. Normalization to global values followed by multivariate analysis using four or five principal components provided optimal discrimination. Connectivity of voxels emerged as an important factor, around 10 connected voxels being optimal for this study.