Accounting for Voxel Neighbourhood Relationship in the SVM
David Hardoon, Janaina Mourão-Miranda, Vanessa Rocha Rego and John Shawe-Taylor
In: 15th Annual Meeting Human Brain Mapping(2008).
In Neuroimage data analysis there are several preprocessing stages that must be applied before any statistical analysis can be done. The sMRI scan preprocessing procedures usually include transformation into standard space, segmentation and smoothing in space using a Gaussian filter. Despite the effectiveness of these preprocessing procedures, any incorrect step may seriously hamper the statistical analysis. Therefore, we propose to incorporate one of the preprocessing steps, namely the smoothing operator, into our neuroimage learning Support Vector Machine (SVM) through the use of a Gaussian prior on the weights. In effect we aim to learn the appropriate smoothing filter while learning the parameterisation for the neuroimage discrimination. We demonstrate our proposed novel technique on smoothed and non-smoothed sMRI in classifying patients and controls.