Abstract:We describe here a classification system based on automatically identified cortical sulci. Multivariate recognition methods are required for the detection of complex brain patterns with a spatial distribution. However, such methods may face the well-known issue of the curse of dimensionality-the risk of overfitting the training dataset in high-dimensional space. We overcame this problem, using a classifier pipeline with one- or two-stage of descriptor selection based on machine-learning methods, followed by a support vector machine classifier or linear discriminant analysis. We compared alternative designs of the pipeline on two different datasets built from the same database corresponding to 151 brains. The first dataset dealt with cortex asymmetry and the second dealt with the effect of the subject's sex. Our system successfully (98%) distinguished between the left and right hemispheres on the basis of sulcal shape (size, depth, etc.). The sex of the subject could be determined with a success rate of 85%. These results highlight the attractiveness of multivariate recognition models combined with appropriate descriptor selection. The sulci selected by the pipeline are consistent with previous whole-brain studies on sex effects and hemispheric asymmetries.

Classification based on cortical folding patterns