Brain MRI foundation models fail on new datasets due to content encoding, not origin signal

A new study published on arXiv reveals that the failure of brain MRI foundation models on unseen datasets is not due to the models encoding the origin of the scans, but rather how the image content itself is encoded. Researchers found that while most models recognize the MRI sequence within a dataset, they struggle to transfer this recognition to new datasets because the content is encoded along different directions. Only models with broad pretraining, like 3DINO and SAM-Med3D, demonstrated consistent sequence encoding for successful transfer, outperforming brain-specific models such as BrainIAC and BrainFM. AI

IMPACT This research highlights a critical limitation in current foundation models for medical imaging, suggesting a need for more robust encoding strategies for reliable cross-dataset generalization.

RANK_REASON The cluster contains a research paper detailing findings about foundation models. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.CV →

• 3DINO
• arXiv
• BrainFM
• Brainiac
• Coral
• foundation model
• INLPTA
• LEACE
• magnetic resonance imaging of the brain
• SAM-Med3D

• paper
• model release

AI-generated summary · Google Gemini · from 1 sources. How we write summaries →