Carlton Chu, Yizhao Ni, Geoffrey Tan, and John Ashburner
University College London & University of Southampton
Kernel Methods for fMRI Pattern Prediction – applications of Relevance Vector Regression and Kernel Ridge Regression
Methods | Poster | Slides
Left to right: Dr. John Ashburner, Carlton Chu, Geoffrey Tan, Yizhao Ni
The statue behind is now believed to a portrayal of Queen Charlotte,
wife of King George III. The background is at the prestigious Queen Square.
The procedure involved realigning the data, detrending with discrete cosine basis, spatial smoothing, and applying multivariate linear and non-linear pattern recognition approaches (i.e. kernel ridge regression, relevance vector regression). The whole fMRI sequences were used, and training was done with the HRF convolved scores. The scans are realigned first, then detrend by a high pass filter. Tissue segmentation was done directly on the EPI scans with SPM5. A gray matter mask was generated to remove other tissue classes, hence only gray matter was used as input features. During the learning and predicting, both kernel ridge regression and relevance vector machine (RVM) were used, and the regularization parameter for ridge regression is learned through cross-validation. Non-linear kernel shows better performance for emotional ratings, and linear kernel performs better for sensory ratings.
Finally, a quadratic programming procedure was carried to perform a constrained deconvolution and reconvolution, which threshold the perdition into reasonable range. The prediction is also smooth with the Gaussian kernel. The advantage of kernel methods is the reduction of computational complexity for high dimensional data as the problem is reduced from number of voxels into number of subjects. There is also a bonus when using a linear kernel, that is the weighting map in the original feature space can be generated. This map indicates the weights of voxels contributing to the predicted rating, and may show interesting functional networks, which might be discovered by mass univariate approaches.
Carlton CHU was born in Taiwan, and later immigrated to New Zealand at age of 16. His undergraduate degree was Computer System Engineering in Auckland University, New Zealand, then he finished my Master of Biomedical Engineering in University of New South Wales, Sydney, Australia. He also worked as part time research assistant in Neuropsychiatric centre in Prince of Wales Hospital, Sydney for half year before starting his PhD in The Wellcome Trust Centre for Neuroimaging, University College London 2005.
His research interests lie in the fields of machine learning and structural image analysis. Carlton's primary supervisor is Dr. John Ashburner. Together they currently collaborate with clinical neurologists and try to apply machine learning methods (mainly kernel methods) on classifying neurological diseases. (Alzheimer's Disease, Huntington's Disease...)
Yizhao Ni - His main research interests are in development and application of machine learning methods, especially kernel based learning algorithms. In the very recent project he focused on the maximize margin methods and structure learning.
Links of Interest: