Unsupervised Joint Alignment Framework for Cross-Site Depression Classification Using fMRI.

Area of Science:

• Neuroimaging
• Machine Learning
• Mental Health Research

Background:

• Major Depressive Disorder (MDD) affects over 264 million globally, causing significant burdens.
• Resting-state functional MRI (rs-fMRI) shows promise for MDD diagnosis.
• Inter-site heterogeneity in rs-fMRI data hinders generalizable diagnostic models.

Purpose of the Study:

• To develop a robust framework for cross-site MDD classification using rs-fMRI.
• To address the challenge of inter-site heterogeneity in machine learning models for MDD detection.

Main Methods:

• Proposed the Unsupervised Joint Alignment (UJA) framework for rs-fMRI data adaptation.
• Utilized a multi-head self-attention module for feature extraction.
• Integrated adversarial domain-wise and class-wise alignment using dual classifiers and sliced Wasserstein distance.

Main Results:

• UJA significantly outperformed existing methods in cross-site MDD classification on the REST-meta-MDD dataset.
• Ablation studies confirmed the effectiveness of the dual alignment strategy.
• Demonstrated the potential for generalizable MDD diagnosis across different imaging sites.

Conclusions:

• UJA offers a promising unsupervised approach for adapting rs-fMRI data across sites.
• The framework enhances the robustness and generalizability of MDD diagnostic models.
• UJA could serve as a valuable tool for clinical decision support in MDD diagnosis.