Classifying Major Depressive Disorder Using Multimodal MRI Data: A Personalized Federated Algorithm

Zhipeng Fan, Jingrui Xu, Jianpo Su, Dewen HuBrain 2025 · DOI

TL;DR

This paper introduces pF-GMCO, a novel personalized federated learning algorithm for diagnosing Major Depressive Disorder (MDD) using multimodal MRI data. By combining adaptive gradient matching and contrastive optimization, it achieves high accuracy (79.07%) across 23 sites while preserving data privacy, offering a more robust and personalized approach than existing federated methods.

What Problem Does This Paper Address?

Diagnosing Major Depressive Disorder (MDD) is challenging, with current methods relying heavily on subjective clinical judgment. While neuroimaging, particularly MRI, shows promise for objective diagnosis, creating robust models is hindered by data heterogeneity across different scanners and acquisition protocols at various research sites. Furthermore, sharing raw MRI data from multiple institutions raises significant privacy, security, and ownership concerns. Federated learning (FL) offers a solution by enabling collaborative model training without direct data sharing, but standard FL methods often struggle with domain shift (variations in data characteristics across sites), leading to suboptimal performance. This paper aims to address these challenges by developing a privacy-preserving, personalized FL framework that effectively handles domain shift and integrates multimodal MRI data for improved MDD classification.

Methods

The core of this work is the Personalized Federated Gradient Matching and Contrastive Optimization (pF-GMCO) algorithm. It introduces two key innovations:

  • Adaptive Gradient Matching: This mechanism adaptively weights the contributions from different participating sites during federated aggregation. It uses cosine similarity to match gradients, prioritizing information from sites with similar data distributions. This helps to mitigate the impact of domain shift.
  • Model Contrastive Optimization: This component encourages personalized model optimization for each site. It uses a contrastive loss function to ensure that the local models retain site-specific characteristics while still benefiting from global knowledge.

Additionally, the paper incorporates Multimodal Compact Bilinear (MCB) pooling to effectively fuse features from structural MRI (sMRI) and functional MRI (fMRI) data (specifically, Amplitude of Low-Frequency Fluctuations (ALFF) and Regional Homogeneity (ReHo)). MCB pooling captures high-order interactions between these modalities, enhancing the discriminative power of the integrated features. The overall approach involves local training at each site, followed by federated aggregation and optimization steps that incorporate both gradient matching and contrastive learning.

Datasets

The study utilizes the Rest-Meta-MDD consortium dataset. This dataset comprises multimodal MRI data from 2293 subjects (1028 MDD patients and 1265 healthy controls) collected from 23 imaging sites. The data includes T1-weighted sMRI and resting-state fMRI. The authors note that the data is publicly available, eliminating the need for ethical approval for their study.

Key Insights and Ideas

  • Personalization is Key: The study strongly advocates for personalized FL models over a single global model, demonstrating that tailoring models to individual site characteristics, while still leveraging shared knowledge, leads to superior performance.
  • Gradient Matching for Domain Shift: The adaptive gradient matching technique proves effective in mitigating the negative impacts of domain shift by intelligently weighting contributions from different sites based on their data similarity.
  • Multimodal Fusion Enhances Accuracy: Combining sMRI and fMRI data through MCB pooling significantly boosts classification performance compared to using single modalities, highlighting the complementary nature of these data types for MDD diagnosis.
  • Privacy-Preserving Diagnosis: The federated learning approach successfully enables collaborative model training without compromising the privacy of sensitive patient data, a critical factor for multisite medical research.
  • Neurobiological Interpretability: The identified discriminative regions of interest (ROIs) align with known MDD pathophysiology, lending credibility to the model's findings and suggesting its potential as a clinical tool.

Key Results

The pF-GMCO algorithm achieved an average classification accuracy of 79.07% on the Rest-Meta-MDD dataset, outperforming existing federated learning strategies and multimodal fusion methods.

  • For single modalities, local models achieved accuracies around 74.55% (sMRI), 58.23% (ALFF), and 60.10% (ReHo).
  • The fused multimodal approach with pF-GMCO reached 79.07% accuracy, significantly improving upon single-modal results and other federated methods.
  • The study also demonstrated the algorithm's robustness against dishonest participants sharing fake information, maintaining high performance.
  • Visualization analysis identified specific brain regions (e.g., temporal cortex, default mode network, frontoparietal network) as discriminative biomarkers for MDD, consistent with existing literature.

Limitations and Future Work

The authors acknowledge several limitations:

  • Backbone Architecture: The use of a single, relatively simple backbone (3D-ResNet-10) for all models might limit its capacity to capture complex neuroimaging features. Future work could explore more specialized architectures.
  • Client Heterogeneity: To manage limited sample sizes at some sites, smaller sites were grouped into larger clients. This aggregation might introduce intraclient heterogeneity and affect local optimization. Future research should focus on reducing this heterogeneity.
  • fMRI Data Acquisition: The rs-fMRI data was acquired at non-standardized times, and diurnal rhythms can affect neuroimaging metrics. Future work should consider recording scan times and participant chronotypes, and developing temporal harmonization methods.
  • Confounding Factors: The study did not account for pharmacological therapy status, a significant confounder in MDD diagnosis. Future studies should include detailed phenotyping of medication histories.
  • Task Specificity: The proposed method was tested solely on MDD diagnosis. Its potential for classifying other neuropsychiatric disorders needs to be explored.

Future work will involve collecting data from more independent medical sites, exploring federated learning in cross-site clinical scenarios, and extending the application to other neuropsychiatric disorders.

Summary updated

Abstract

Evaluated on the Rest-Meta-MDD dataset with 2293 subjects from 23 sites, pF-GMCO achieved accuracy of 79.07%, demonstrating superior performance and interpretability. This work provides an effective and privacy-aware framework for multisite MDD diagnosis using federated learning.
BibTeX
@inproceedings{fan2025classifying,
  author = {Zhipeng Fan and Jingrui Xu and Jianpo Su and Dewen Hu},
  title = {{Classifying Major Depressive Disorder Using Multimodal MRI Data: A Personalized Federated Algorithm}},
  booktitle = {Brain},
  year = {2025},
  doi = {10.3390/brainsci15101081},
}
Figure