Skip to content

submission-guide.md

Latest commit

 

History

History
121 lines (82 loc) · 4.48 KB

submission-guide.md

File metadata and controls

121 lines (82 loc) · 4.48 KB

Submission Guide

This guide explains how to submit predictions at each phase of the BIC-MAC Challenge.

Phase Overview

Phase Period What you submit What we return
Validation May 15 – Aug 15 Zip of NIfTI predictions uploaded to Codabench All metrics on the 4 validation subjects
Dry Run May 15 – Aug 15 Docker container via email CT metrics on the 4 validation subjects (or error logs if the container failed)
Final Test June 15 - Aug 15 Docker container via email Full evaluation on the unseen test set (September 1)

Validation and Dry Run run concurrently throughout the challenge. Use them to iterate on your model before the final deadline. There is no limit on submissions during either phase.

Phase 1: Validation — NIfTI Upload

Submit your predictions directly as NIfTI files. No Docker container needed.

What to submit

Run your model on the 4 validation subjects (you have both features/ and recon/ for these) and produce predictions:

  1. Pseudo-CT (ct.nii.gz) — run your model on features/
  2. Reconstructed PET (pet.nii.gz, optional) — run the reconstruction pipeline on your pseudo-CT using the provided Docker image (see reconstruction.md)

If you only submit ct.nii.gz, you will receive CT metrics only. Submitting both unlocks all four metrics.

Output requirements

  • NIfTI format (.nii.gz)
  • Same shape and affine as features/nacpet.nii.gz
  • CT values in Hounsfield units

Zip structure

submission.zip
├── sub-004/
│   ├── ct.nii.gz
│   └── pet.nii.gz   # optional
├── sub-009/
│   ├── ct.nii.gz
│   └── pet.nii.gz   # optional
├── sub-010/
│   ├── ct.nii.gz
│   └── pet.nii.gz   # optional
└── sub-018/
    ├── ct.nii.gz
    └── pet.nii.gz   # optional

Upload to the Codabench competition page.

Phase 2: Dry Run — Container Check

The dry run verifies that your Docker container runs correctly on organizer hardware before the final deadline. Submit as early as possible to leave time to fix issues.

We run your container on the 4 validation subjects and return either:

  • CT metrics — your container ran successfully and produced valid pseudo-CTs
  • Error logs — the container failed, with details of what went wrong

See docker-packaging.md for how to build and test your container locally before submitting.

How to submit

Email bic-mac-challenge@github.io with subject line [DRY-RUN] <TeamName> and include:

  • Team name, Docker image name and tag
  • A link to your image using one of the options below

Option A — Docker Hub (preferred):

docker tag my-model:latest <dockerhub-username>/my-model:latest
docker push <dockerhub-username>/my-model:latest

Send us the full image name (e.g. myteam/my-model:latest).

Option B — Compressed archive via file sharing:

docker save my-model:latest | gzip > my-model.tar.gz

Upload my-model.tar.gz to Google Drive, Dropbox, or similar and share the download link.

Dry Run submissions are limited to two per month per team.

Phase 3: Final Test

Submit your Docker container by August 15, 2026. The container does not need to be the same as the one used for the dry run — you can continue to improve your model right up to the deadline.

We will:

  1. Run your container on each unseen test subject
  2. Run the full reconstruction pipeline on each pseudo-CT
  3. Evaluate all metrics against ground-truth CT and PET

Results and winner announcements: September 1, 2026.

How to submit

Same as the dry run — email your container to bic-mac-challenge@github.io with subject [FINAL] <TeamName>, using Docker Hub or a compressed archive with a file sharing link. Make sure to also include a link to a short methedology paper describing your approach. This methedology paper must be uploaded to a public repository and it is a requirement to be considered eligeble for prizes.

Hardware Constraints (Phases 2 & 3)

Resource Specification
GPU 1× NVIDIA A40
CPU 2× Intel Xeon Gold 6346 @ 3.10 GHz
RAM 128 GB
Wall-clock time per subject 5 minutes
Network access None (--network none)

All weights and dependencies must be baked into the image. No downloads at inference time.