They used a two-stage training strategy.
MRI foundation model aligns anatomy and abnormalities across 3D volumes
Trained on 50,000 studies, A4-MR outperforms prior models on abnormality detection, segmentation, and localization without anatomical prompts.
Big Tech
Noel D'souza · Zhijian Yang · Nishank Singla · Pranay Doshi · Krisztian Koos · Fei Tan · +11 more
GE Healthcare · Amazon
Research Digest··2 min read
The authors introduce A4-MR, a 3D vision-language foundation model for MRI trained on approximately 50,000 studies.
Why this paper
From Amazon and GE Healthcare
In one line
A 3D MRI vision-language model that aligns image-text representations at the anatomical level and separates healthy from abnormal anatomy outperforms existing MRI foundation models.
What we could check
- ·No code link found
- ·No weights link found
- ·No dataset link found
- ·No compute details found
- ·No stated limitations found
- ·No benchmark numbers found
Observed from the paper text and links we have. Absence here means we did not find it, not that it does not exist.
§