Neuroscience

Learning Sparse Latent Predictive Foundation Model for Multimodal Neuroimaging

Haoxu Huang, Long Chen, Jingyun Chen, Jinu Hyun, James Ryan Loftus, Kara Melmed, Daniel Orringer, Jennifer Frontera, Seena Dehkharghani, Arjun Masurkar, Narges Razavian

Featured June 20, 2026

Original Paper

This analysis was generated by SciGrove. Upload your own PDFs or enter a DOI — and get the same AI breakdown on any paper.

Get started

AI-generated analysis — This is SciGrove's AI interpretation of the paper, not peer-reviewed content. Always refer to the original paper.

Simply

A new AI model called Neuro-JEPA learns to understand different types of brain MRI scans together by predicting missing parts of images, using special 'expert' networks, and focusing only on the brain tissue, making it much better at finding diseases than older methods.

In depth

The paper introduces Neuro-JEPA, a sparse multimodal neuroimaging foundation model that combines a latent predictive objective (JEPA) with a Mixture-of-Experts (MoE) architecture. This integration, along with specialized multiscale masking and a foreground-aware loss, enables the model to learn unified representations across diverse MRI sequences, achieving robust and consistent performance in clinical neuroimaging tasks.

Key Takeaways

1
Neuro-JEPA integrates a latent predictive objective (JEPA) with a Mixture-of-Experts (MoE) architecture for sparse, multimodal neuroimaging representation learning.
2
The model introduces multiscale masking and a foreground-aware L1 loss to optimize pretraining for 3D neuroimaging data, improving anatomical prediction and robustness.
3
Neuro-JEPA consistently outperforms existing neuroimaging foundation models and CNN baselines across a wide range of unimodal and multimodal clinical tasks, demonstrating superior generalization and label efficiency.

Conceptual Flow

HIGH LEVEL

Methodology: Learning Unified Brain Features

The model learns by predicting hidden parts of brain scans using specialized 'expert' networks, focusing only on important brain areas.

Brain Scan (T1w)

Brain Scan (T2w)

Brain Scan (FLAIR)

Combine & Learn Patterns

Unified Brain Understanding

Results: Superior Performance in Disease Detection

This new way of learning helps the model perform much better and more consistently across many different brain disease tasks compared to other AI models.

Old AI Models

Neuro-JEPA Model

Compare Performance

Better Disease Detection

More Reliable Results