Welcome to nnsyn! ( 🏆 1st place in MICCAI-SynthRAD2025 MR-CT synthesis challenge)

This repo holds the code and docker, which won 1st place in MR-to-CT synthesis task in MICCAI SynthRAD2025 challenge.

• SynthRAD2025 Task1 Final Leaderboard - 🥇 1st place [15-Aug-2025]
• SynthRAD2025 Task2 Post-challenge Leaderboard - 🥇 1st place [post-challenge, 17-Aug-2025]

✨ What is nnsyn? Self-configured framework for medical image synthesis

In this project, we would like to produce a user-friendly, mask-supported, extendable framework for medical image synthesis. We incorporated new CT preprocessing, new network architectures, new loss functions, and new evaluation metrics for image synthesis tasks.

🌟 Feature highlights:

• Oneliner preprocessing
• Oneliner training (support masked loss, support MedNext)
• Oneliner inference
• Train a dedicated segmentation branch for perception loss
• Optional advanced experiment tracking with AIM
• Support on multimodal imaging inputs (ongoing)

🚀 Installation:

git clone git@github.com:aehrc/nnsyn.git
cd nnsyn
pip install -e .

📄 Quick start

First, export environment variables :

export nnsyn_origin_dataset = "path_to/nnsyn_origin/synthrad2025_task1_mri2ct_AB"
export nnUNet_raw="path_to/nnUNet_raw"
export nnUNet_preprocessed="path_to/nnUNet_preprocessed"
export nnUNet_results="path_to/nnUNet_results"

Organise your data into "nnsyn_origin_dataset". The "MASKS" folder contains the body contour, while the 'LABELS' folder contains segmentation labels. An example of dataset.json in example. Currently, data needs to be convert to .mha.

DATA_STRUCT:
|-- nnsyn_origin
|   |-- synthrad2025_task1_mri2ct_AB
|       |-- INPUT_IMAGES
|           |-- PATIENT_1_0001.mha
|       |-- TARGET_IMAGES
|           |-- PATIENT_1_0001.mha
|       |-- MASKS (optional)
|           |-- PATIENT_1.mha
|       |-- LABELS (optional)
|           |-- PATIENT_1.mha
|           |-- dataset.json 
|-- nnUNet_raw
|   |-- DatasetXXX_YYY
|-- nnUNet_preprocessed
|   |-- DatasetXXX_YYY
|-- nnUNet_results
|   |-- DatasetXXX_YYY

Plan experiments and preprocess for the synthesis model.

nnsyn_plan_and_preprocess -d 960 -c 3d_fullres -pl nnUNetPlannerResEncL -p nnUNetResEncUNetLPlans  --preprocessing_input MR --preprocessing_target CT 

(For loss_map) Prepare dataset and preprocess for the segmentation model. The plan will be transfered from synthesis model (960) to segmentation model (961).

nnsyn_plan_and_preprocess_seg -d 960 -dseg 961 -c 3d_fullres -p nnUNetResEncUNetLPlans

(For loss_map) Train the segmentation model for perception loss. We first switch to github segmentation branch (nnunetv2), train the segmentation model, and then switch back to the github synthesis branch (main).

git switch nnunetv2
nnUNetv2_train 961 3d_fullres 0 -tr nnUNetTrainer -p nnUNetResEncUNetLPlans_Dataset960 --c
git switch main

Train the synthesis network with Masked Anatomical Perception (map) loss:

nnsyn_train 960 3d_fullres 0 -tr nnUNetTrainer_nnsyn_loss_map -p nnUNetResEncUNetLPlans

Inference :

nnsyn_predict -d 960 -i INPUT_PATH -o OUTPUT_PATH -m MASK_PATH -c 3d_fullres -p nnUNetResEncUNetLPlans -tr nnUNetTrainer_nnsyn_loss_map -f 0

🤝 Credit

This project was build upon nnUNet_translation, nnUNet-v2, and TriALS. All awesome stuff. Please do not hesitate to check them out.

• nnUNet_translation: https://github.com/Phyrise/nnUNet_translation - Paper1 - Paper2
• nnUNet-v2: https://github.com/MIC-DKFZ/nnUNet - Paper
• TriALS: https://github.com/xmed-lab/TriALS

ℹ️ Docker & huggingface space

Please find the instructions to reproduce the docker image for SynthRAD2025 task1 and task2 at https://huggingface.co/aehrc/synthrad2025_docker.

Also, we provided a demo at huggingface space. Because only cpu resources are available for demo, it would be a bit slow (5 min/volume). On gpu, the inference time is 9 seconds/volume.