Disease classification: A probabilistic approach

We describe a probabilistic technique for separating two populations whereby analysis is performed on affine-invariant representations of each patient. The method begins by converting each voxel from a high-dimensional diffusion weighted signal to a low-dimensional diffusion tensor representation. Three orthogonal measures that capture different aspects of the local tissue are derived from the tensor representation to form a feature vector. From these feature vectors, we form a probabilistic representation of each patient. This representation is affine invariant, thus obviating the need for registration of the images. We then use a Parzen window classifier to estimate the likelihood of a new patient belonging to either population. To demonstrate the technique, we apply it to the analysis of 22 first-episode schizophrenic patients and 20 normal control subjects. With leave-many-out cross validation, we find a detection rate of 90.91% (10% false positives).