STS-Skill

Your entry point for all MICCAI STS Challenge tasks. This skill routes you to the right workflow based on what you want to do.

How to Use

When the user mentions STS challenge, teeth segmentation, Pre-Task, or any competition task, follow this routing logic:

Step 1: Identify User Intent

Parse the user's message for these keywords and route accordingly:

Step 2: Load the Guide

Once you identify the task, read the corresponding GUIDE.md file. The guide contains step-by-step instructions that you should follow to help the user complete their task.

Step 3: Execute

Follow the guide's instructions. For scripts, use the appropriate tool (Bash) to run them. For code tasks, write/modify code as instructed.

Menu (show when intent is unclear)

When the user's intent is unclear, present this menu:

🏆 STS Challenge Skill - What would you like to do?

📋 2026 Tasks:
  1. Pre-Task    - 2D segmentation (UNet, ~1 hour)
  2. Task 1      - Metal artifact removal
  3. Task 2      - [Task description]
  4. Task 3      - [Task description]

📚 Historical:
  5. STS 2025    - 3D CBCT pulp segmentation
  6. STS 2024    - Instance segmentation
  7. STS 2023    - 2D teeth segmentation (semi-supervised)

🔧 Tools:
  8. Download data
  9. Evaluate predictions
  10. Prepare submission

Just tell me the number or describe what you need!

Shared Workflows

These patterns are used across multiple tasks. Reference them from the guides.

Environment Check

Before any task, verify the user's environment:

python --version          # Need 3.8+
python -c "import torch; print(f'PyTorch {torch.__version__}')"
python -c "import torch; print(f'CUDA available: {torch.cuda.is_available()}')"

Data Download

Use scripts/download_data.py to download data from the fastest available source:

python scripts/download_data.py --task <task-name> --output ./data

Supported tasks: pretask-2026

Local Evaluation

Use scripts/evaluate.py to check prediction quality before submitting:

python scripts/evaluate.py --pred <predicted_masks_dir> --gt <ground_truth_dir>

Submission Packaging

Use scripts/prepare_submit.py to create a Codabench-compatible zip:

python scripts/prepare_submit.py --masks <masks_dir> --task <task-name>

Data Sources

Pre-Task data is available from:

Competition Links

References

For deeper understanding, read:

• references/algorithms.md — UNet, semi-supervised learning, loss functions
• references/competition-history.md — STS 2023-2026 overview
• references/faq.md — Common questions and troubleshooting