Learning Sparse Latent Predictive Foundation Model for Multimodal Neuroimaging
Researchers have developed Neuro-JEPA, a novel foundation model designed to learn unified representations from multimodal brain MRI scans. This model utilizes a sparse latent predictive objective and a Mixture-of-Experts architecture to process T1w, T2w, and FLAIR imaging sequences. Pretrained on over 1.5 million scans, Neuro-JEPA demonstrated superior and more consistent performance across 25 clinical and research tasks compared to existing neuroimaging foundation models and a CNN baseline. AI
IMPACT Establishes a scalable framework for multimodal neuroimaging representation learning, potentially improving diagnostic accuracy and research insights.