unicorn-baseline 1.4.4

Baseline inference algorithm for the UNICORN Challenge

UNICORN Baseline 🦄

Welcome to the official baseline repository for the UNICORN challenge!
This repository provides reference implementations and tools for tackling a wide range of vision, language, and vision-language tasks in computational pathology and radiology.

This baseline uses the following publicly available foundation models:

• Virchow (link to publication)
• PRISM (link to publication)
• MRSegmentator (link to publication)
• CT-FM: Whole Body Segmetation (link to publication)
• phi4
• BioGPT (link to publication)
• opus-mt-en-nl

🚀 Quickstart

System requirements: Linux-based OS (e.g., Ubuntu 22.04) with Python 3.10+ and Docker installed.
We provide scripts to automate the local testing process using public few-shot data from Zenodo.

1. Clone the Repository

git clone https://github.com/DIAGNijmegen/unicorn_baseline.git
cd unicorn_baseline

2. Download Model Weights

⚠️ Access Required
Some of the models used in the baseline are gated.
You need to have requested and been granted access to be able to download them from Hugging Face.

./download_weights.sh

3. Build the Docker Container

./do_build.sh

4. Perform test run(s)

Make sure to always take the latest version of the data on Zenodo.

• Single Task: Downloads and prepares data for a single task, then runs the docker on one case.

  ./run_task.sh "https://zenodo.org/records/15315589/files/Task01_classifying_he_prostate_biopsies_into_isup_scores.zip"
  

• All Tasks: Runs the docker on all supported UNICORN tasks, sequentially.

  ./run_all_tasks.sh
  

• Targeted Test Run: Run the docker on a specific case folder.

  ./do_test_run.sh path/to/case/folder [docker_image_tag]
  

5. Save the Container for Submission

./do_save.sh

📝 Input & Output Interfaces

• Input: Each task provides a unicorn-task-description.json describing the required inputs and metadata. See example-data/ for sample files and structure.
• Output: The baseline generates standardized output files (e.g., image-neural-representation.json, patch-neural-representation.json) as required by the challenge.

📜 License

This project is licensed under the Apache License 2.0.