🧠 FOMO25 Challenge Codebase

This repository contains the official baseline code for the FOMO25 Challenge, which investigates the few-shot generalization properties of foundation models for brain MRI data analysis.

Please note: This codebase will be continually refined, so check back occasionally to get the latest updates.

Data and Model Download Links

• FOMO-60K is available for download via Huggingface.
• Baseline models pretrained on FOMO-60K are available for download via Huggingface.

🔍 Challenge Overview

The FOMO25 Challenge seeks to advance the field of medical image analysis by evaluating the same pretrained models across multiple downstream tasks. Participants will first pretrain on a large unlabelled dataset and then evaluate their models on three clinical, multi-vendor, and multi-center datasets. For more information on the challenge, please visit the FOMO25 Challenge website.

⏱️ Timeline

• April 1: Challenge Opens
• April 7: Code Release
• April 15: Sanity-Check Code Access
• June 15: Validation Leaderboard and Final Submission Pipeline Opens
• August 20: Challenge Submission Deadline
• September 27, 2025: Results Release at MICCAI 2025

📋 Tasks and Data

This codebase supports three tasks:

1. Task 1: Infarct Detection - Binary classification of brain infarcts
2. Task 2: Meningioma Segmentation - Binary segmentation of brain meningiomas
3. Task 3: Brain Age Regression - Predicting patient age from MRI scans

Data for the challenge includes:

• Pretraining Data: 11,187 subjects, 13,900 sessions, 60,529 scans
• Finetuning Data: Limited few-shot data (~20-200 cases per task)
• Evaluation Data: Unseen data for final assessment

🛠️ Requirements

Install required dependencies:

# Install basic dependencies
pip install -e .

# For development
pip install -e ".[dev]"

# For testing
pip install -e ".[test]"

# For all dependencies
pip install -e ".[dev,test]"

Data Preparation

While the data included in this challenge is already preprocessed (co-registered, transposed to RAS orientation and defaced/skull-stripped), to run this code, one needs to further preprocess with the following highly opinionated preprocessing steps.

This "Opinionated Preprocessing" can be done in the following way

Yucca Preprocess Pretraining Data (required)

For preprocessing the pretraining (FOMO60K) data:

python src/data/fomo-60k/preprocess.py --in_path=/path/to/raw/pretrain/data --out_path=/path/to/output/preprocessed/data

This will:

1. Store each tensor in numpy format for easy loading.
2. Treat each scan as a separate datapoint which can be sampled iid.
3. Crop to the minimum bounding box.
4. Z-normalize on a per-volume level.
5. Resample to isotropic (1mm, 1mm, 1mm) spacing.

Yucca Preprocess Finetuning Data (required)

For preprocessing the finetuning data for tasks 1-3:

python src/data/preprocess/run_preprocessing.py --taskid=1 --source_path=/path/to/raw/finetuning/data

Replace --taskid=1 with --taskid=2 or --taskid=3 for the other tasks.

This will:

1. Assemble each session into a single 4D tensor and store it as a numpy array for easy loading.
2. Crop to the minimum bounding box.
3. Z-normalize on a per-volume level.
4. Resample to isotropic (1mm, 1mm, 1mm) spacing.

Pretraining

To pretrain a model using the AMAES (Augmented Masked Autoencoder) framework:

python src/pretrain.py \
    --save_dir=/path/to/save/models \
    --pretrain_data_dir=/path/to/preprocessed/pretrain/data \
    --model_name=unet_b_lw_dec \
    --patch_size=96 \
    --batch_size=2 \
    --epochs=100 \
    --warmup_epochs=5 \
    --num_workers=64 \
    --augmentation_preset=all

Key pretraining parameters:

• --model_name: Supported models include unet_b_lw_dec, unet_xl_lw_dec, etc.
• --patch_size: Size of 3D patches (must be divisible by 8)
• --mask_patch_size: Size of masking unit for MAE (default is 4)
• --mask_ratio: Ratio of patches to mask (default is 0.6)
• --augmentation_preset: Choose from all, basic, or none

Finetuning

To finetune a pretrained model on one of the three tasks:

python src/finetune.py \
    --data_dir=/path/to/preprocessed/data \
    --save_dir=/path/to/save/finetuned/models \
    --pretrained_weights_path=/path/to/pretrained/checkpoint.pth \
    --model_name=unet_b \
    --patch_size=96 \
    --taskid=1 \
    --batch_size=2 \
    --epochs=500 \
    --train_batches_per_epoch=100 \
    --augmentation_preset=basic

Key finetuning parameters:

• --taskid: Task ID (1: Infarct Detection, 2: Meningioma Segmentation, 3: Brain Age Regression)
• --model_name: Must match the architecture of the pretrained checkpoint without lw_dec, so if you pretrained with unet_xl_lw_dec, this should be unet_xl.
• --pretrained_weights_path: Path to the pretrained model checkpoint
• --augmentation_preset: Choose from all, basic, or none

💾 Model Checkpoints

Pre-trained model checkpoints coming soon...

Note: When finetuning a specific checkpoint, make sure to use the matching model architecture. For example, to finetune unet_xl_lw_dec_fullaug.pth, use --model_name=unet_xl.

💻 Hardware Requirements

The reference implementation was pretrained on 2xH100 GPUs with 80GB of memory. Depending on your hardware, you may need to adjust batch sizes and patch sizes accordingly.

📝 Submission Guidelines

Participants can compete in two tracks:

1. Methods Track 🧪: Only using the provided FOMO60K dataset for pretraining
2. Open Track 🌐: Using any data for pretraining (private data allowed)

For both tracks, additional data is not allowed for finetuning. More details about the submission process can be found on the FOMO25 Challenge website.

📚 Citation

If you use this code, please cite:

@article{munk2024amaes,
  title={AMAES: Augmented Masked Autoencoder Pretraining on Public Brain MRI Data for 3D-Native Segmentation},
  author={Munk, Asbjørn and Ambsdorf, Jakob and Llambias, Sebastian and Nielsen, Mads},
  journal={MICCAI Workshop on Advancing Data Solutions in Medical Imaging AI (ADSMI 2024)},
  year={2024}
}

@article{llambias2024yucca,
  title={Yucca: A deep learning framework for medical image analysis},
  author={Llambias, Sebastian N{\o}rgaard and Machnio, Julia and Munk, Asbj{\o}rn and Ambsdorf, Jakob and Nielsen, Mads and Ghazi, Mostafa Mehdipour},
  journal={arXiv preprint arXiv:2407.19888},
  year={2024}
}