Cortical Surface Thickness as a Classifier: Boosting for Autism Classification

Abstract

Abstract. We study the problem of classifying an autistic group from controls using structural image data alone, a task that requires a clinical interview with a psychologist. Because of the highly convoluted brain surface topology, feature extraction poses the first obstacle. A clinically relevant measure called the cortical thickness has shown promise but yields a rather challenging learning problem -where the dimensionality of the distribution is extremely large and the training set is small. By observing that each point on the brain cortical surface may be treated as a "hypothesis", we propose a new algorithm for LPBoosting (with truncated neighborhoods) for this problem. In addition to learning a high quality classifier, our model incorporates topological priors into the classification framework directly -that two neighboring points on the cortical surface (hypothesis pairs) must have similar discriminative qualities. As a result, we obtain not just a label {+1, −1} for test items, but also an indication of the "discriminative regions" on the cortical surface. We discuss the formulation and present interesting experimental results.

Individual subject classification for Alzheimer's disease based on incremental learning using a spatial frequency representation of cortical thickness data

“…Therefore, their vertices with the same index correspond to each other. The cortical thickness at each vertex is defined as the distance between the two surfaces at the vertex (Singh et al (2008)). However, in order to establish the correspondences between the vertices of different subjects, the cortical thickness data of each subject needs to be represented in a common space.…”

Candidate Biomarkers in Children with Autism Spectrum Disorder: A Review of MRI Studies

“…Among ASD patients, the SVM was first applied in adults. Singh et al developed a diagnostic model generated by the LPboost-based algorithm to distinguish autistic children from controls, based on voxel-wise cortical thickness and ~40,000 points for each individual; they reported 89% classification accuracy based on cross-validation [171]. Ecker et al (2010) applied SVM classifiers to investigate the predictive value of whole-brain structural volumetric changes in ASD, and obtained 81% classification accuracy based on cross-validation [172].…”

Topological Characterization of Signal in Brain Images Using Min-Max Diagrams

“…We incrementally decreased the size of the training set (up to 35%) and found that the algorithm still gives more than 96% accuracy. A simple comparison with 90% accuracy reported in [8] that uses the same data suggests that the improvements in accuracy comes primarily from our min-max representation.…”

“…We evaluated linear and Gaussian weighted kernels (using Bhattacharya distance between the two PDFs [5]) and found that the accuracy results were quite similar. To perform our evaluations relative to existing techniques, we used data shared with us by the authors in [8]. We summarize our results next.…”