Source-free collaborative domain adaptation via multi-perspective feature enrichment for functional MRI analysis.

Area of Science:

• Neuroimaging
• Machine Learning
• Medical Informatics

Background:

• Resting-state functional MRI (rs-fMRI) is vital for multi-site neurological disorder research.
• Cross-site data heterogeneity, stemming from scanner and protocol variations, poses a significant challenge.
• Current domain adaptation techniques often necessitate access to source domain data, raising privacy and storage concerns.

Purpose of the Study:

• To propose a novel source-free collaborative domain adaptation (SCDA) framework.
• To address the limitations of existing methods by eliminating the need for source data access.
• To enhance the analysis of neurological disorders using heterogeneous rs-fMRI data.

Main Methods:

• Developed a source-free collaborative domain adaptation (SCDA) framework.
• Introduced a multi-perspective feature enrichment (MFE) method to leverage target fMRI data from multiple views.
• Employed an unsupervised pretraining strategy on a large dataset (3,806 unlabeled fMRIs) and initialized MFE with a pretrained source model for efficient knowledge transfer.

Main Results:

• Demonstrated the efficacy of the SCDA framework in reducing rs-fMRI data heterogeneity.
• Achieved successful cross-scanner and cross-study prediction performance.
• Validated the method on three public and one private dataset.

Conclusions:

• The proposed SCDA framework offers a viable solution for domain adaptation in multi-site rs-fMRI studies.
• It effectively mitigates cross-site data heterogeneity without requiring access to source data.
• This approach facilitates more robust and generalizable analysis of neurological disorders.