Addressing multi-site functional MRI heterogeneity through dual-expert collaborative learning for brain disease...

Area of Science:

• Neuroscience
• Medical Imaging
• Machine Learning

Background:

• Multi-site resting-state functional MRI (rs-fMRI) studies for major depressive disorder (MDD) identification face challenges due to inter-site heterogeneity.
• Existing models often fail to generalize across multiple target domains because of variations in scanners and scanning protocols.

Purpose of the Study:

• To propose a Dual-Expert fMRI Harmonization (DFH) framework for automated MDD diagnosis.
• To mitigate data distribution differences across multiple unlabeled target domains by leveraging a single labeled source domain.

Main Methods:

• The DFH framework employs a domain-generic student model and two domain-specific teacher/expert models.
• Joint training is performed using a deep collaborative learning module for knowledge distillation.
• The approach simultaneously utilizes data from one labeled source domain and two unlabeled target domains.

Main Results:

• The proposed DFH framework yields a student model with enhanced generalizability, adaptable to unseen target domains.
• Comprehensive experiments on 836 subjects across 3 sites demonstrate the superiority of the DFH method.
• The identified discriminative brain functional connectivities show potential as biomarkers for fMRI-related MDD diagnosis.

Conclusions:

• The DFH framework represents a significant advancement in multi-target fMRI harmonization for MDD diagnosis.
• The derived generalizable model can potentially be applied to the analysis of other brain diseases.
• This study is among the first to address multi-target fMRI harmonization for MDD identification.