Multiple Kernel Learning Classification of Lupus Disease from Structural MRI

• Activity Number: 256
• Type: Contributed
• Date/Time: Monday, August 1, 2011 : 2:00 PM to 3:50 PM
• Sponsor: Section on Statistical Learning and Data Mining
• Author(s): Cen Guo*+
• Companies: University of Michigan
• Keywords: MRI ; Multiple Kernel Learning

Abstract

Structural Magnetic Resonance imaging (MRI) was used as in the clinical trails extensively in recent years. Classification analysis of anatomical MRI can help in diagnosis of disease. Traditional classification analysis using svm method is either based on whole-brain image with tens of thousands of variables or a particular region of interest which needs extra information about the disease to specify. In this work, we proposed a hierarchical method to reduce the dimension and select significant regions automatically. The first step is to train a svm model for every small region across the whole brain. In the second step, a multiple kernel learning scheme is applied to significant regions to further improve the performance and select important features simultaneously. A real data analysis of Lupus disease shows that this new method can outperform the traditional one step svm method.