Abstract
Autism is a developmental disorder characterized by social deficits, impaired communication, and restricted and repetitive patterns of behavior (American Psychiatry Association, 2000). Various neuropathological studies of autism have revealed abnormalities in several brain regions. Increased head size was the first observed characteristic in children with autism. According to the published studies, different anatomical structures of the brain have been identified as being involved in the abnormal neurodevelopment associated with autism. Classical neuropathological studies as well as MRI structural findings are consistent with respect to some brain structures, while observations with respect to other structures have differed among studies. This lack of consistency may be due to the sample size as well as to the failure to account for significant confounding factors (e. g., age, sex, IQ, handedness). However, there is an increasing agreement from structural imaging studies on the abnormal anatomy of the white matter (WM) in autistic brains. In addition, deficits in the size of the corpus callosum (CC) and its subregions in patients with autism relative to controls are well established. In this work, we aim at using the reported abnormalities of the WM and the CC, in order to devise robust classification methods of autistic vs. normal subjects through analyzes of their respective MRIs. To overcome the limitations and shortcomings of the volumetric studies, our analysis is based on shape descriptions and geometric models. A novel technique is used to compute the 3D distance map as a shape descriptor of the WM. The distribution of this distance map is then used as a statistical feature that allows discrimination between the two groups. Furthermore, we use our newly proposed nonrigid registration technique based on scale space and curve evolution theories to devise a new classification approach by analyzing the deformation fields (DF) generated from registering CCs onto each others. For each group (autistic and control), we pick a number of segmented CC datasets, one of which is chosen as reference and the remaining ones registered to this reference using our registration method. The generated DFs are averaged, and their distributions are computed to represent the changes of the magnitudes of each group. Given a subject to be classified, we register its MR dataset, once to the chosen control reference and then to the chosen autistic reference. The two corresponding DFs are statistically compared to the two average DFs, representing each class, in order to indicate the class to which the tested subject belong. The accuracy of our techniques was tested on postmortem and in-vivo brain MR data. The results are very promising and show that, contrary to traditional methods, the proposed techniques are less sensitive to age and volume effects.