BIC-MAC Challenge Codebase

Big Cross-Modal Attenuation Correction — synthesize pseudo-CT from multi-modal PET/MRI input to enable CT-less PET reconstruction.

🏆 Challenge website | 🗂️ Dataset | 🏆 CodaBench submission & leaderboard

---

Updates

• May 15, 2018: NIfTI validation and leaderboards now live - see submission instructions at https://www.codabench.org/competitions/12555
• April 8, 2026 (13:00 GMT+2): [BUG FIX] A bug in the Carney HU-to-MU conversion formula has been fixed. Please make sure to run docker pull ghcr.io/bic-mac-challenge/recon:latest and git pull for updated recon and evaluation logic. The 8 ground truth PET train images have been updated on huggingface.
• April 7, 2026: [NEW DATA POLICY] The use of public datasets for pretraining and other use-cases is now allowed under certain conditions. Please see docs/rules.md for details.

Table of Contents

• Overview
• Documentation
• Getting Started
• Dataset Structure
• Pseudo-CT Baseline
• Reconstruction
• Evaluation
• Submission

---

🧠 Overview

Your algorithm receives the files under features/ for each subject and must output a pseudo-CT volume as a NIfTI file in Hounsfield units (HU). Predictions are evaluated two ways (see Evaluation for metric definitions):

1. CT accuracy — Predicted pseudo-CT is compared directly against the ground-truth CT
2. PET accuracy — Predicted pseudo-CT is fed into the reconstruction pipeline to produce an attenuation-corrected PET image, which is then compared against the ground-truth PET

Note that no PET reconstruction experience is needed to participate in the challenge, and the main purpose of the reconstruction is to enable clinically meaningful metrics.

---

📚 Documentation

Guide	Description
Rules	All rules for participating teams including training data policy, and pretraining policy
Data Background	Details dat acquisition, preprocessing, and alignment of modalities
Reconstruction Pipeline	How the pseudo-CT is turned into an AC-PET image; how to run it locally
Submission Guide	Validation, dry-run, and final submission phases explained
Docker Packaging	How to containerize your model, with baseline as a worked example
Tips & FAQ	Common questions, pitfalls, and practical advice

---

🚀 Getting Started

Requirements: Python 3.12, uv, Docker

uv sync

The src/ directory contains three components:

src/
├── baseline/       # Baseline pseudo-CT model (patch-based 3D UNet)
├── evaluation/     # Metric defintinitions and scripts
└── recon/          # PET reconstruction script (and Docker)

---

🗂️ Dataset structure

The dataset comprises 99 subject-unique studies, with 20 reserved for testing and the remaining 79 available on huggingface and split as follows:

Split	Subjects	Contents
train/ (full)	8	features/ + ct-label/ + recon/ + pet-label/
train/ (no recon)	67	features/ + ct-label/
val/	4	features/ + recon/

All train subjects have CT labels, but due to the size of the sinograms, only 8 include the recon/ and pet-label/ folders needed for closed loop reconstruction. Validation subjects have sinogram recon/ data but no labels — submit predicted pseudo-CTs and reconstructed PETs to Codabench to get live leaderboard metrics during the challenge. The train subjects with recon/ data are: sub-000, sub-001, sub-002, sub-005, sub-006, sub-008, sub-013, sub-014.

All images except those in pet-label are resampled to the label CT image (tensor size: 512x512x531, voxel size 1.52x1.52,2.00mm^3). NIfTI images are structured in four folders per subject.

• features/ All the files you can use as input to your pseudo-CT model during inference.
• ct-label/ The CT target (ct.nii.gz) and segmentations for evaluation.
• pet-label/ The PET target (pet.nii.gz) and segmentations for evaluation.
• recon/ Sinograms and metadata for PET reconstructions.

train/
└── sub-000/
    ├── features/                          # generative model inputs
    │   ├── nacpet.nii.gz                  # non-attenuation-corrected PET. 
    │   ├── topogram.nii.gz                # 2D scout X-ray
    │   ├── mri_chunk_{0-3}_{in/out}_phase.nii.gz    # MRI chunk (0-3), in- and out-phase
    │   ├── mri_combined_{in/out}_phase.nii.gz  # stitched whole-body MRI, in- and out-phase
    │   ├── mri_face_mask.nii.gz           # binary anonymization mask
    │   └── metadata.json                  # {sex, age, height, weight}
    ├── ct-label/                          # ground-truth CT
    │   ├── ct.nii.gz                      # CT in Hounsfield Units (reference)
    │   ├── body_seg.nii.gz                # TotalSegmentator body seg.
    │   ├── organ_seg.nii.gz               # TotalSegmentator organ seg.
    │   └── prediction_mask.nii.gz         # Within-body voxels (face excluded).
    ├── recon/                             # sinogram data for reconstruction
    │   ├── mult_nac_rd85.hs/.s            # multiplicative sinogram (NAC)
    │   ├── add_nac_rd85.hs/.s             # additive sinogram (NAC)
    │   ├── prompts_rd85.hs/.s             # raw sinogram 
    │   ├── offset.json                    # bed position and gantry offset
    │   ├── ct_face_and_bed.nii.gz         # inverse of prediction_mask.nii.gz
    │   └── face_and_bed_mask.nii.gz       # binary face + scanner bed mask
    └── pet-label/                         # ground-truth PET
        ├── pet.nii.gz                     # CT-attenuation-corrected PET (reference)
        ├── body_seg.nii.gz                # TotalSegmentator body seg. in PET space
        ├── tissue_seg.nii.gz              # TotalSegmentator tissue seg. in PET space
        └── organ_seg.nii.gz               # TotalSegmentator organ seg. in PET space

---

📦 Pseudo-CT Baseline (src/baseline/)

A simple patch-based 3D UNet that predicts pseudo-CT from NAC-PET only.

Python usage:

# Example:
python src/baseline/predict.py --features_dir data/sub-000/features/ --output_ct results/sub-000/ct.nii.gz

Docker usage:

docker pull ghcr.io/bic-mac-challenge/baseline:latest

docker run --rm \
  --gpus all \
  -v /path/to/sub-000/features:/data/features:ro \
  -v /path/to/output:/data/output \
  ghcr.io/bic-mac-challenge/baseline:latest

The predicted CT is written to /data/output/ct.nii.gz. All weights and dependencies are baked into the image, and the same is expected for your final docker image submission.

You can re-train the baseline by running train.py and containerize it by running docker build -t my-baseline . (from inside the src/baseline folder)

---

⚙️ Reconstruction (src/recon/)

Converts a CT (ground-truth or pseudo-CT) and PET sinograms into an attenuation-corrected PET image using STIR. See docs/reconstruction.md for pipeline details and local usage instructions.

docker pull ghcr.io/bic-mac-challenge/recon:latest

docker run --rm \
  -v /path/to/sub-000/recon:/data/recon \
  -v /path/to/ct.nii.gz:/data/ct/ct.nii.gz \
  -v /path/to/output:/data/output \
  ghcr.io/bic-mac-challenge/recon:latest

The reconstructed PET is written to /data/output/pet.nii.gz.

Warning

Running reconstruction requires ~20 GB of RAM and takes 20–120 minutes depending on CPU speed. The intermediates/ folder uses ~50 GB of additional disk space — consider deleting it after a successful reconstruction.

---

📊 Evaluation (src/evaluation/)

Five metrics compare predicted PET and CT outputs against the ground truth:

Metric	Modality	Description	Region
Whole-body SUV MAE	PET	Mean absolute error in standardised uptake value (SUV = activity × weight / total dose)	Body mask, excluding ±4 cm around liver
Brain Outlier Score	PET	AUC of fraction of brain voxels within relative error thresholds (5%, 10%, 15%) Ladefoged 2017 (fig. 8)	Brain
Organ Bias	PET	Mean absolute relative error of mean SUV in 8 organs: brain, liver, spleen, heart, pancreas, muscle, adipose, extremities	TotalSegmentator organ labels
CT MU MAE	CT	Mean absolute error of attenuation coefficients (μ at 511 keV) between predicted and ground-truth CT after HU→μ conversion	Body mask, excluding ±4 cm axial slices at top of liver
TAC Bias	Dynamic PET	Absolute relative error of the integral of time-activity-curves (TACs) for the aorta and selected brain regions. NOTE: Metric is computed only for the final test set due to the size of the dynamic sinograms.	Brain regions and aorta

Evaluate a single subject:

python src/evaluation/eval_subject.py \
  --subject_dir <subject_dir> \
  --pred_pet <pred_pet.nii.gz> \
  --pred_ct <pred_ct.nii.gz>

--pred_pet and --pred_ct are both optional — omit either to skip PET or CT metrics. Note: Brain Outlier Score is a dataset-level metric and requires multiple subjects (see below).

Evaluate multiple subjects:

python src/evaluation/eval_dataset.py \
  --dataset_dir <dataset_dir> \
  --pred_dir <predictions_dir>

<predictions_dir> must contain one sub-folder per subject with consistent contents — either CT only, PET only, or both:

predictions_dir/
├── sub-000/
│   ├── ct.nii.gz        
│   └── pet.nii.gz       
├── sub-001/
│   ├── ct.nii.gz
│   └── pet.nii.gz
└── ...

---

📬 Submission

Phase	What you submit
Validation	Zip of NIfTI predictions for the 4 val/ subjects (CT + optional PET) uploaded to Codabench — you run prediction and reconstruction locally using the provided data and tools
Dry Run	Docker container emailed to us — we run it on the 4 val/ subjects on our hardware and return CT metrics, or error logs if the container failed
Final Test	Docker container emailed to us — we run prediction, reconstruction, and full evaluation on the unseen test set

Validation and Dry Run open May 15. See docs/submission-guide.md for full instructions, and docs/docker-packaging.md for how to build and test your container.