Detection of white matter abnormalities in MR brain images for diagnosis of autism in children

This paper introduces a novel computer-aided diagnosis (CAD) system for the diagnosis of autism from magnetic resonance (MR) brain images of children. The proposed framework has two main components. First, cerebral white matter (CWM) is segmented from the brain volume using an adaptive shape model that is built from a set of co-aligned training images. A higher-order Markov Gibbs random field (MGRF) spatial interaction model is then integrated with an intensity model to account for data in homogeneities. Secondly, CWM meshes are reconstructed and a curvature-based analysis is applied in order to differentiate between autistic and control brains. From the reconstructed CWM meshes, we estimated 3 shape features (curvedness, sharpness, and mean curvature) that could geometrically describe CWM folds. The statistical analysis conducted on 20 autistic and control brains using binomial regression on raw features as well as on their principal components reveals that the three curvature-based measures could be used as discriminatory features between autistic and control brains. Moreover, the accuracy of the diagnostic results using K-nearest neighbor (KNN) classifier is 95%. These results show the promise of the proposed technique as a supplement to current diagnostic instruments (e.g., Autism Diagnostic Interview, Revised (ADI-R)).