Comparative study of SVM methods combined with voxel selection for object category classification on fMRI data

Area of Science:

• Neuroimaging
• Machine Learning
• Biomedical Data Analysis

Background:

• Support Vector Machines (SVM) are established for analyzing functional MRI (fMRI) data.
• Prior research indicated no significant advantage of non-linear (polynomial kernel) SVM over linear SVM.
• This study introduces a comparative analysis of linear SVM and a more effective non-linear SVM utilizing a radial basis function (RBF) kernel.

Purpose of the Study:

• To investigate the combined performance of linear and RBF SVM for fMRI classification.
• To evaluate the impact of various voxel selection methods on classification accuracy and computational efficiency.
• To determine optimal SVM and voxel selection strategies for fMRI data analysis.

Main Methods:

• Employed six distinct voxel selection techniques for fMRI data.
• Utilized both linear and RBF kernels within SVM classifiers for 4-category object classification.
• Compared the overall performance of different voxel selection and classification method combinations.

Main Results:

• Voxel selection significantly influenced classification accuracy.
• RBF SVM demonstrated superior performance over linear SVM in lower-dimensional feature spaces.
• Linear SVM outperformed RBF SVM in higher-dimensional feature spaces.
• Optimal accuracy and efficiency were achieved by linear SVM with more voxels or RBF SVM with a reduced voxel set post-PCA.

Conclusions:

• This research offers empirical insights into linear and RBF SVM performance for fMRI classification coupled with voxel selection.
• For maximizing classification accuracy, RBF SVM with selected voxels or linear SVM with more voxels are recommended.
• For prioritizing computational speed, RBF SVM with a reduced voxel set (principal components) is the preferred approach.