Autistic Subjects Classification Based on PET Images

Background: Most of recent brain imaging studies of autism are based on methods that look for abnormalities at a voxel level ignoring distributed pattern abnormalities. A second limitation of those methods is that they only enable group analysis (eg.: control vs. patient) of populations.

Objectives: Conversely to voxel-based method we propose a method that simultaneously considers several brain regions taking into account their informative interrelations in order to improve the sensitivity of abnormalities detection. Moreover, such method enables individual analysis, which offers new perspectives like the evaluation the efficiency of a treatment on a single subject.

Methods: Rest cerebral blood flow (CBF) images of 45 (5 to 12 years) children with primary autistic disorder and 13 nonautistic (5 to 15 years) children with idiopathic mental retardation were measured with a positron emission tomography (PET) camera. We propose a multi-stages pattern recognition method. Early stages aim to identify autism discriminant regions. Later stage selects the most discriminant combination of those regions to build a autism vs. control classifier. Finally, the classifier is applied to new individual images and the prediction (autistic, control) is compared with the true clinical outcome to evaluate the accuracy of the individual classifier.

Results: The classification system identified two regions: (i) right superior temporal sulcus (hypoperfusion in autism) (ii) The left postcentral gyrus (hyperperfusion in autism). The validation of the method yielded to 91% (41/45, P=0.04) of correct classification of autistic subjects and 77% (10/13, P=5.10^-9) for mental retardation subjects.

Conclusions: Efficient individual detection of rest CBF abnormalities in children with autism may be obtained from PET images without any a priori manually defined ROIs.