Isolating Nonlinear Independent Sources in fMRI with $\beta$-TCVAE Models

2026-05-19 · Source: stat.ML updates on arXiv.org · Field: Science & Research — Artificial Intelligence & Machine Learning, Life Sciences & Biology, Health & Medical Research · Depth: Expert, long

Summary

A pilot study adapted and modified the $\beta$-TCVAE (Total Correlation Variational Autoencoder) framework for nonlinear source disentanglement in fMRI data. This approach aims to separate mixed spatial and temporal brain signals into interpretable components, addressing limitations of traditional linear independent component analysis (ICA) in capturing complex brain dynamics. The subject-conditioned variational autoencoder model incorporates subject-specific embeddings and optimizes a decomposed ELBO objective, including reconstruction, mutual information, total correlation, and dimension-wise KL terms. Evaluated on the Human Connectome Project (HCP) dataset, the $\beta$-TCVAE successfully recovered biologically relevant nonlinear spatial components, such as the default mode network (DMN) and orbitofrontal cortex (OFC), demonstrating more spatially coherent and complete representations compared to InfoMax ICA. The framework also captured multi-network coupling patterns, suggesting improved preservation of interactions between distributed brain networks.

Key takeaway

For AI Scientists and Research Scientists working with neuroimaging data, this study demonstrates that adapting the $\beta$-TCVAE framework can yield more robust and interpretable functional brain networks than traditional ICA. You should consider integrating nonlinear representation learning methods like $\beta$-TCVAE into your fMRI analysis pipelines, especially when seeking to capture complex, multi-network interactions and improve the biological plausibility of your extracted components.

Key insights

$\beta$-TCVAE effectively disentangles nonlinear fMRI data into biologically meaningful and spatially coherent brain networks.

Principles

• Nonlinear models capture complex brain dynamics better than linear ICA.
• Disentangled representations improve interpretability of fMRI components.
• Subject-conditioned models account for inter-subject variability.

Method

The $\beta$-TCVAE adapts a subject-conditioned variational autoencoder, optimizing an ELBO decomposition to disentangle fMRI signals into spatial and temporal components, using a learned subject embedding and Adam optimizer.

In practice

• Apply $\beta$-TCVAE for advanced fMRI network identification.
• Use subject embeddings to model inter-subject variability.
• Consider $\beta$-TCVAE for multi-network co-activation analysis.

Topics

• fMRI Analysis
• Nonlinear Source Disentanglement
• beta-TCVAE
• Independent Component Analysis
• Intrinsic Connectivity Networks

Best for: AI Scientist, Research Scientist

Related on AIssential

• Fusion Learning from Dynamic Functional Connectivity: Combining the Amplitude and Phase of fMRI Signals to Identify Brain Disorders — Combining fMRI signal amplitude and phase via fusion learning improves brain disorder detection.
• BrainDINO: A Brain MRI Foundation Model for Generalizable Clinical Representation Learning — Large-scale slice-wise self-supervised learning yields a unified, transferable brain MRI representation for diverse neuroimaging tasks.
• Feature selection leads to divergent neurobiological interpretations of brain-based machine learning biomarkers — Discarded brain features can predict phenotypes with accuracy comparable to top-ranked features, yielding divergent neurobiological interpretations.
• Pan-FM: A Pan-Organ Foundation Model with Saliency-Guided Masking for Missing Robustness — Pan-FM is a pan-organ foundation model robust to missing data via saliency-guided masking.
• Continual Learning for fMRI-Based Brain Disorder Diagnosis via Functional Connectivity Matrices Generative Replay — A continual learning framework uses generative replay and knowledge distillation to diagnose brain disorders from sequential fMRI data.
• Learning Sparse Latent Predictive Foundation Model for Multimodal Neuroimaging — Neuro-JEPA is a sparse multimodal neuroimaging foundation model using a latent predictive objective and Mixture-of-Experts.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by stat.ML updates on arXiv.org.