Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

BrainCSD: A Hierarchical Consistency-Driven MoE Foundation Model for Unified Connectome Synthesis and Multitask Brain Trait Prediction

Xiongri Shen, Jiaqi Wang, Yi Zhong, Zhenxi Song, Leilei Zhao, Liling Li, Yichen Wei, Lingyan Liang, Shuqiang Wang, Baiying Lei, Demao Deng, Zhiguo Zhang

TL;DR

BrainCSD addresses the bottlenecks of preprocessing-heavy connectome analysis and missing modalities by introducing a hierarchical mixture-of-experts foundation model that jointly synthesizes FC and SC from uni-modal inputs while enabling downstream diagnosis and trait prediction. The three-stage design—ROI Activation MoE, Encoding Activation MoE, and Network-Aware Finetuning—enforces activation consistency and neuroanatomical priors through contrastive learning and network-specific refinements, yielding high-fidelity connectome reconstructions and robust task performance. Across multi-site datasets and tasks, BrainCSD achieves state-of-the-art synthesis metrics, strong MCI/HC/AD and PD diagnosis under missing modalities, and accurate brain-age and MMSE predictions, demonstrating practical potential for clinical deployment. The work provides a scalable, biologically grounded framework that bypasses extensive preprocessing and offers reliable, interpretable connectome synthesis and decoding in real-world settings.

Abstract

Functional and structural connectivity (FC/SC) are key multimodal biomarkers for brain analysis, yet their clinical utility is hindered by costly acquisition, complex preprocessing, and frequent missing modalities. Existing foundation models either process single modalities or lack explicit mechanisms for cross-modal and cross-scale consistency. We propose BrainCSD, a hierarchical mixture-of-experts (MoE) foundation model that jointly synthesizes FC/SC biomarkers and supports downstream decoding tasks (diagnosis and prediction). BrainCSD features three neuroanatomically grounded components: (1) a ROI-specific MoE that aligns regional activations from canonical networks (e.g., DMN, FPN) with a global atlas via contrastive consistency; (2) a Encoding-Activation MOE that models dynamic cross-time/gradient dependencies in fMRI/dMRI; and (3) a network-aware refinement MoE that enforces structural priors and symmetry at individual and population levels. Evaluated on the datasets under complete and missing-modality settings, BrainCSD achieves SOTA results: 95.6\% accuracy for MCI vs. CN classification without FC, low synthesis error (FC RMSE: 0.038; SC RMSE: 0.006), brain age prediction (MAE: 4.04 years), and MMSE score estimation (MAE: 1.72 points). Code is available in \href{https://github.com/SXR3015/BrainCSD}{BrainCSD}

BrainCSD: A Hierarchical Consistency-Driven MoE Foundation Model for Unified Connectome Synthesis and Multitask Brain Trait Prediction

TL;DR

BrainCSD addresses the bottlenecks of preprocessing-heavy connectome analysis and missing modalities by introducing a hierarchical mixture-of-experts foundation model that jointly synthesizes FC and SC from uni-modal inputs while enabling downstream diagnosis and trait prediction. The three-stage design—ROI Activation MoE, Encoding Activation MoE, and Network-Aware Finetuning—enforces activation consistency and neuroanatomical priors through contrastive learning and network-specific refinements, yielding high-fidelity connectome reconstructions and robust task performance. Across multi-site datasets and tasks, BrainCSD achieves state-of-the-art synthesis metrics, strong MCI/HC/AD and PD diagnosis under missing modalities, and accurate brain-age and MMSE predictions, demonstrating practical potential for clinical deployment. The work provides a scalable, biologically grounded framework that bypasses extensive preprocessing and offers reliable, interpretable connectome synthesis and decoding in real-world settings.

Abstract

Functional and structural connectivity (FC/SC) are key multimodal biomarkers for brain analysis, yet their clinical utility is hindered by costly acquisition, complex preprocessing, and frequent missing modalities. Existing foundation models either process single modalities or lack explicit mechanisms for cross-modal and cross-scale consistency. We propose BrainCSD, a hierarchical mixture-of-experts (MoE) foundation model that jointly synthesizes FC/SC biomarkers and supports downstream decoding tasks (diagnosis and prediction). BrainCSD features three neuroanatomically grounded components: (1) a ROI-specific MoE that aligns regional activations from canonical networks (e.g., DMN, FPN) with a global atlas via contrastive consistency; (2) a Encoding-Activation MOE that models dynamic cross-time/gradient dependencies in fMRI/dMRI; and (3) a network-aware refinement MoE that enforces structural priors and symmetry at individual and population levels. Evaluated on the datasets under complete and missing-modality settings, BrainCSD achieves SOTA results: 95.6\% accuracy for MCI vs. CN classification without FC, low synthesis error (FC RMSE: 0.038; SC RMSE: 0.006), brain age prediction (MAE: 4.04 years), and MMSE score estimation (MAE: 1.72 points). Code is available in \href{https://github.com/SXR3015/BrainCSD}{BrainCSD}

Paper Structure

This paper contains 19 sections, 18 equations, 6 figures, 11 tables.

Table of Contents

  1. INTRODUCTION
  2. Related Work
  3. Methods
  4. ROI Activation MoE
  5. Encoding Activation MoE
  6. Network-Aware MoE Finetuning
  7. Loss Functions
  8. Experiment
  9. Datasets and Downstream Evaluation
  10. Implementation Details
  11. Stage 1: ROI Activation.
  12. Stage 2: Encoding Activation.
  13. Stage 3: Network-Aware Finetuning.
  14. Why This Design?
  15. Results
  16. ...and 4 more sections

Figures (6)

  • Figure 1: Hierarchical consistency-driven (consistency-D) MoE for connectome synthesis and prediction under missing modalities, via ROI (Network $\leftrightarrow$ Atlas) → Encoding (time[t]/gradient [g] point $\leftrightarrow$ Whole series)→ Neuroanatomy-informed MOE.
  • Figure 2: BrainCSD: ROI → Encoding → Network refinement for hierarchical activation-consistent connectome synthesis
  • Figure 3: Network-Level Masking Strategy. Regions from six canonical networks are selectively masked during training to enforce region-specific feature learning.
  • Figure 4: Network-Aware Finetuning. Each network is refined by a dedicated MoE expert, ensuring anatomical and diagnostic consistency.
  • Figure 5: Group-Level Consistency. Connectivity features are regularized to be consistent within clinical/demographic subgroups, enhancing diagnostic utility.
  • ...and 1 more figures