Identifying Spatial Differences in Multiple Sclerosis Subtypes via a Log Gaussian Cox Process

• Activity Number: 372
• Type: Invited
• Date/Time: Tuesday, July 31, 2012 : 2:00 PM to 3:50 PM
• Sponsor: Biometrics Section
• Author(s): Jian Kang and Timothy D. Johnson*+ and Thomas E. Nichols
• Companies: Emory University and University of Michigan and University of Warwick
• Keywords: Spatial point process ; Bayesian Model ; Multiple Sclerosis ; Prediction ; log-Gaussian Cox process ; Parallel computing

Abstract

Multiple Sclerosis (MS) is a disease in which the myelin sheaths surrounding the axons of the brain and spinal cord are damaged. This leads to dymelination and scarring of the white matter tracks in the brain. There are five main subtypes of MS: 1) clinically isolated syndrom 2) relapsing/remitting, 3) secondary progressive, 4) progressive relapsing and 5) primary progressive. Researchers are interested in whether there are differences in both location and intensity of MS lesions among these subtypes. To answer this question we model the locations of MS lesions by a log Gaussian Cox process (LGCP). Estimation is performed within the Bayesian framework. We estimate the LGCP on the entire 3D image via the Fast Fourier Transform (FFT). Computational efficiency is realized by parallelizing the FFT algorithm and performing the calculations on a graphical processing unit. The resulting models can be used to predict a new patient's subtype given his/her lesion pattern by invoking Bayes' Theorem. We show, via an importance sampling method, that we achieve excellent cross-validated prediction results.