Researchers have discovered that frozen foundation models, when used to represent brain MRI data, primarily encode the site where the MRI was acquired rather than anatomical or clinical information. This effect was observed across different cohorts, encoders, and network depths, with site being highly decodable even from randomly initialized models or raw images. While methods like ComBat can remove this site fingerprint, it comes at the cost of entangling site and anatomical information within a shared subspace. The study recommends auditing foundation models for site attribution and provides an open toolkit for this purpose. AI
IMPACT Highlights potential biases in foundation models used for medical imaging, necessitating careful auditing for site-specific artifacts.
RANK_REASON The cluster contains an academic paper detailing research findings on foundation models. [lever_c_demoted from research: ic=1 ai=1.0]
- ABIDE-II
- Abide In Me
- ComBat
- foundation model
- Hugging Face
- magnetic resonance imaging of the brain
- residual neural network
- Swin Transformer
- Vít
AI-generated summary · Google Gemini · from 1 sources. How we write summaries →
