MAMBA-BASED RESIDUAL GENERATIVE ADVERSARIAL NETWORK FOR FUNCTIONAL CONNECTIVITY HARMONIZATION DURING INFANCY.

Area of Science:

• Neuroimaging
• Machine Learning
• Data Harmonization

Background:

• Site effects pose a significant challenge in multi-site neuroimaging studies.
• Existing methods for harmonizing resting-state functional magnetic resonance imaging (rs-fMRI) functional connectivity (FC) are underdeveloped.
• Statistical models are inadequate for capturing the complex, nonlinear relationships in FC data.

Purpose of the Study:

• To develop a novel deep learning method for harmonizing multi-site rs-fMRI functional connectivity (FC).
• To address the limitations of current statistical models in handling nonlinear FC data.
• To improve the reliability and comparability of neuroimaging data across different research sites.

Main Methods:

• Development of a Mamba-based Residual Generative Adversarial Network (MR-GAN).
• Leveraging Mamba Blocks to identify FC-specific sequential patterns.
• Integration with a multi-task residual GAN for harmonizing multi-site FC data.

Main Results:

• The MR-GAN method demonstrated superior performance in harmonizing multi-site FC data.
• Experiments were conducted on 939 infant rs-fMRI scans from four distinct sites.
• The proposed method effectively mitigated site effects while preserving biological characteristics.

Conclusions:

• MR-GAN offers a flexible and effective deep learning solution for harmonizing multi-site rs-fMRI FC data.
• The Mamba Block integration enhances the ability to capture complex sequential patterns in FC.
• This advancement is critical for advancing multi-site neuroimaging research and analysis.