Meeting Banner
Abstract #3125

Symmetric and Multi-Scale Features for Automatic Segmentation of Multiple Sclerosis Lesions Using Pattern Classification

Marco Battaglini1, Nicola De Stefano1, Mark Jenkinson2

1Quantitative Neuroimaging Laboratory, University of Siena, Siena, Italy; 2Clinical Neurology, FMRIB Centre, University of Oxford, Oxford, Oxon, United Kingdom

IIn order to develop a fully automated segmentation tool for MS lesions we explore using novel input features with two pattern classification methods (Neural Networks and Random Forests). Results show a statistically significant improvement in DICE by using the novel multi-scale and symmetry features with both classifiers. To be useful for clinical trials we use multi-centre real clinical data, segmented by different manual raters, which makes this challenging. Nonetheless, we still achieve DICE results consistent with state-of-the-art methods, without requiring costly pruning of the Neural Networks, complicated post-processing, or having to apply any exclusion criteria to the images.