MuVi: Official repository of "Single Image Test-Time Adaptation via Multi-View Co-Training" In MICCAI 2025.
This pipeline is built on top of the nnU-Net inference pipeline, extending the nnUNetPredictor class.
📂 Repository Structure
├── LICENSE # Main license for this repo
├── NOTICE # Notices for reused third-party code
├── README.md # You’re reading it :)
│
├── images/ # Figures for README/docs
|
└── scripts/ # All experiment & training scripts
│
├── compute_metrics.py # Evaluation metrics (e.g. Dice, Hausdorff, ASSD)
├── infer_with_test_time_adaptation.py # 🌟 Main entry point for running inference w/ MuVi and other methods
├── run.sh # Example bash script to run experiments
├── utils.py # Shared utility functions
│
├── test_time_training_methods/ # Implementations of different TTA methods
│ ├── bnadapt/ # BNAdapt method
│ │ ├── LICENSE
│ │ ├── NOTICE
│ │ ├── bn.py # BNAdapt core implementation
│ │
│ ├── intent/ # InTent method
│ │ ├── LICENSE
│ │ ├── intent.py # InTent core implementation
│ │
│ ├── memo/ # MEMO method
│ │ ├── LICENSE
│ │ ├── memo.py # MEMO core implementation
│ │ └── test_time_augmentation.py # Augmentation logic for MEMO
│ │
│ ├── muvi/ # 🌟 Our proposed MuVi method
│ │ ├── config.yaml # 🌟 MuVi hyperparameters & configs
│ │ └── muvi_trainer.py # 🌟 MuVi training & test-time adaptation logic
│ │
│ └── tent/ # Tent method (modified the original repo to add PTN and 2D BatchNorm --> 3D BatchNorm)
│ ├── LICENSE
│ ├── README.md # Original Tent docs
│ ├── cfgs/
│ │ ├── norm.yaml
│ │ ├── source.yaml
│ │ └── tent.yaml
│ ├── cifar10c.py
│ ├── conf.py
│ ├── norm.py # PTN core implementation
│ ├── tent.py # Tent core implementation
│ └── requirements.txt
Follow the official nnU-Net installation instructions:
👉 nnU-Net GitHub
⚠️ Our pipeline is built on nnU-Net’s inference pipeline (nnUNetPredictorclass).
Currently, only one-fold inference is supported (not five-fold).
git clone https://github.com/your-username/muvi-tta.git
cd muvi-tta/scripts
📊 Normalization: To switch between InstanceNorm and BatchNorm, edit norm_op in plans.json file in the respective nnUNet_results folder.
📦 Checkpoints: Download the baseline and supervised checkpoints from here and add it to the respective nnUNet_results folder.
python infer_with_test_time_adaptation.py \
--input /path/to/images \
--output /path/to/save/results \
--model /path/to/nnunet_trained_model \
--method muvi
python compute_metrics.py \
--predictions /path/to/save/results \
--ground_truth /path/to/gt_labels
run.sh compiles the commands above!
This repository is currently being developed. The remaining tasks include:
[ ] Add preprocessing pipeline
[ ] Add checkpoints to google drive
[ ] Add data partitioning information
[ ] Add additional results
[ ] Verify that MuVi works with the current version of nnUNet
[ ] Create requirements.txt file
If you use MuVi in your work, we’d love it if you gave us a shout-out by citing our paper!
@misc{joshi2025singleimagetesttimeadaptation,
title={Single Image Test-Time Adaptation via Multi-View Co-Training},
author={Smriti Joshi and Richard Osuala and Lidia Garrucho and Kaisar Kushibar and Dimitri Kessler and Oliver Diaz and Karim Lekadir},
year={2025},
eprint={2506.23705},
archivePrefix={arXiv},
primaryClass={cs.CV},
url={https://arxiv.org/abs/2506.23705},
}
This work builds upon nnUNet repository and adapts several test time adaptation methods to this framework:
We acknowledge the authors of these works for releasing their code.
This repository was developed at the BCN-AIM, Universitat de Barcelona, Spain.
We gratefully acknowledge support from RadioVal (European research and innovation programme grant agreement No 101057699).
