UNICORN
Unified benchmark for imaging in computational pathology, radiology, and natural language
UNICORN is a public benchmark for evaluating multimodal foundation models across 20 tasks spanning computational pathology, radiology, and medical language processing — all within a single unified framework.
The challenge follows a one-to-many design: a single model is evaluated across all tasks using few-shot adaptation only, with no large-scale task-specific training. This tests true generalization across modalities and domains.
Tasks
- 6 pathology vision tasks
- 5 radiology vision tasks
- 8 language and vision-language tasks
Repositories
-
unicorn_baseline— official baseline with Docker submission framework and pre-integrated foundation models (Virchow, PRISM, CT-FM, Phi-4, …) -
unicorn_eval— evaluation toolkit with an extensible adaptor library for converting frozen features into predictions
The challenge ran as part of MICCAI 2025. See the challenge page and the preprint for full details.