Application of Randomized Singular Value Decomposition Techniques to the Analysis of Imaging Data Using Partial Least Squares

• Activity Number: 288
• Type: Contributed
• Date/Time: Tuesday, August 8, 2006 : 10:30 AM to 12:20 PM
• Sponsor: Biometrics Section
• Author(s): Scott Ziolko*+ and Lisa Weissfeld and Chester Mathis and William Klunk and Steven DeKosky and Julie Price
• Companies: University of Pittsburgh and University of Pittsburgh and University of Pittsburgh and University of Pittsburgh and University of Pittsburgh and University of Pittsburgh
• Keywords: covariance ; PLS ; brain imaging ; PET imaging ; Alzheimer's disease

Abstract

The application of PLS to PET imaging data requires the computation of the singular value decomposition (SVD) of the covariance matrix of the image data with the covariate data. This approach is computationally intensive in simpler settings where a single modality, such as PET, is used and no longer feasible when using an additional imaging modality in place of the covariate data. Building on previous work (Ziolko et al., 2005), we examine the application of randomized SVD techniques (Drinea, Drineas, Huggins, 2005) to PLS. We compare the use of weighted selection techniques, random sampling techniques, and deletion of voxels with small t-values to reduce the dimensionality of the problem. We focus on the results obtained from PLS using these different approaches.