Abstract
This paper describes a segment-based asymmetry feature detection approach for three-dimensional positron emission tomography (PET) brain images to automatically extract pathological lesions. The method consists of three stages: preprocessing, segmentation, and asymmetry detection. The method was tested on simulation and clinical data sets and a per-pixel asymmetry feature detection is experimentally compared with our per-segment approach and the per-segment method is shown to produce fewer false positives and better demarcation in the PET data examples presented.