Nonlinear KCCA in fMRI activation analysis: Self-supervised optimization and robust back-reconstruction.

Area of Science:

• Neuroimaging
• Machine Learning
• Statistical Analysis

Background:

• Nonlinear kernels have been extended to Kernel Canonical Correlation Analysis (KCCA) for flexible modeling of complex relationships.
• Existing methods for KCCA may have limitations in parameter optimization, inverse mapping accuracy, and kernel selection strategies.

Purpose of the Study:

• To propose three key enhancements to nonlinear KCCA.
• To improve the modeling of complex relationships in functional Magnetic Resonance Imaging (fMRI) data.
• To enhance the accuracy, robustness, and overall performance of nonlinear KCCA.

Main Methods:

• Refined parameter optimization using a subject-wise criterion inspired by self-supervised learning to mitigate overfitting.
• Introduced an improved back-reconstruction (inverse mapping) method for higher accuracy and robustness.
• Investigated and validated a kernel selection strategy based on convergence behavior, activation accuracy, data augmentation robustness, and eigendecomposition.

Main Results:

• The proposed enhancements demonstrate consistent improvements across multiple performance metrics on simulated and task-based fMRI datasets.
• The improved back-reconstruction method shows higher accuracy and robustness compared to existing voxel-importance estimation techniques.
• The validated kernel selection strategy proves effective in enhancing the performance of the nonlinear KCCA framework.

Conclusions:

• The proposed framework offers significant advancements in nonlinear Kernel Canonical Correlation Analysis.
• These enhancements lead to more accurate and robust modeling of complex relationships in fMRI data.
• The study validates the effectiveness of the proposed subject-wise optimization, improved inverse mapping, and kernel selection strategies.