A Self-Supervised Voxel Shuffling Framework for Kernel-Based fMRI activation detection

Motivation: Kernel-based methods are efficient for global feature extraction and activation detection in fMRI analysis. Properly defining the kernel mapping function requires supervised training, which can easily overfit due to high dimensionality, even for linear kernels. Goal(s): Design a data augmentation framework to generate supervisory examples without requiring ground truth or new datasets. The augmented data should preserve the original information, and by maximizing the similarity, we should get more accurate results. Approach: By shuffling the voxels based on their activation properties. Results: In two simulated and two real fMRI datasets, our method can select the best hyperparameters and generate more accurate activation patterns. Impact: The proposed augmentation method effectively finds the optimal kernel mapping and mitigates overfitting in activation detection without relying on spatial ground truth information. This could be used in real fMRI data and broaden the potential for modeling more complex relationships.