Dynamic Gradient Modulation for Enhanced Resting-State FMRI Denoising

Resting-state functional MRI (rs-fMRI), a pivotal tool for probing the brain's spontaneous activity, is frequently marred by noise from non-neuronal sources, obscuring the subtle signals crucial for understanding neural dynamics. Traditional denoising methods, e.g., FIX, depend on manually engineered features, limiting their ability to handle complex noise. While deep learning approaches have shown promise in overcoming these limitations, they often suffer imbalanced optimization problems when differentiating noise from signal based on spatial or temporal features. To navigate these challenges, we introduce the Multi-Level Gradient Modulation (MLGM) framework, a novel approach designed to harmonize the optimization process across spatial and temporal dimensions. MLGM dynamically adjusts gradients to support effective learning across both spatial and temporal features, preventing dominance by either modality. The framework incorporates a Multi-Scale Temporal Feature Extraction technique, which captures a broad range of temporal dynamics while minimizing redundancy. Evaluations across two infant datasets demonstrate that MLGM significantly improves noise removal, achieving superior accuracy and robustness compared to existing methods. fMRI, fMRI Denoising, Gradient Modulation, Multi-scale temporal feature