This is a PyTorch official implementation of paper Med-Tuning: A New Parameter-Efficient Tuning Framework for Medical Volumetric Segmentation. Our paper has been accepted by Medical Imaging with Deep Learning 2024 (MIDL2024).

The ``pre-training then fine-tuning (FT)" paradigm is widely adopted to boost the model performance of deep learning-based methods for medical volumetric segmentation. However, conventional full FT incurs high computational and memory costs. Thus, it is of increasing importance to fine-tune pre-trained models for medical volumetric segmentation tasks in a both effective and parameter-efficient manner. In this paper, we introduce a new framework named Med-Tuning to realize parameter-efficient tuning (PET) for medical volumetric segmentation task and an efficient plug-and-play module named Med-Adapter for task-specific feature extraction. With a small number of tuned parameters, our framework enhances the 2D baselines's precision on segmentation tasks, which are pre-trained on natural images. Extensive experiments on three benchmark datasets (CT and MRI modalities) show that our method achieves better results than previous PET methods on volumetric segmentation tasks. Compared to full FT, Med-Tuning reduces the fine-tuned model parameters by up to 4x, with even better segmentation performance. Our project webpage is at this URL.

Basic requirements:

• Ubuntu 20.04.1
• CUDA version==11.2
• Python==3.8

You need to run the following commands to create a new environment.

conda create -n medt python==3.8
conda activate medt
cd ../Med-Tuning-Official-main
pip install requirements.txt

Download the BraTS2019 datasets and organize them:

- Datasets
    - BraTS2019
        - Train # Official train set, 335 samples
            - HGG
                - BraTS19_xx0
                    - BraTS19_xx0_t2.nii.gz
                    - BraTS19_xx0_t1ce.nii.gz
                    - BraTS19_xx0_t1.nii.gz
                    - BraTS19_xx0_seg.nii.gz
                    - BraTS19_xx0_flair.nii.gz
                    - BraTS19_xx0_data_f32b0.pkl
                - BraTS19_xx1
                ...
                
            - LGG
                - BraTS19_xx0
                    - BraTS19_xx0_t2.nii.gz
                    - BraTS19_xx0_t1ce.nii.gz
                    - BraTS19_xx0_t1.nii.gz
                    - BraTS19_xx0_seg.nii.gz
                    - BraTS19_xx0_flair.nii.gz
                    - BraTS19_xx0_data_f32b0.pkl
                - BraTS19_xx1
                ...

            - train.txt # Contains all official train samples. Each line is is a relative path of one sample dir, eg. HGG/BraTS19_xx0
            - train_0.txt
            - train_1.txt
            - train_2.txt
            - train_3.txt
            - train_4.txt # Contains 4/5 official train samples for 5-fold training.
            - valid_0.txt
            - valid_1.txt
            - valid_2.txt
            - valid_3.txt
            - valid_4.txt # Contains 1/5 official train samples for 5-fold validation.
        - Test # Official test set, 125 samples
            - BraTS19_xx0
                - BraTS19_xx0_t2.nii.gz
                - BraTS19_xx0_t1ce.nii.gz
                - BraTS19_xx0_t1.nii.gz
                - BraTS19_xx0_flair.nii.gz
                - BraTS19_xx0_data_f32b0.pkl
            - BraTS19_xx1
            ...
            - valid.txt # Contains all official test samples. Each line is is a relative path of one sample dir, eg. BraTS19_xx0

Run the following commands to download the supervised pre-trained checkpoint of Swin-T and ViT-B/16.

cd ../Med-Tuning-Official-main/preckpt/
wget https://storage.googleapis.com/vit_models/imagenet21k/ViT-B_16.npz
wget https://github.com/SwinTransformer/storage/releases/download/v1.0.0/swin_tiny_patch4_window7_224.pth

Then change the corresponding path in configs/config.py.

First change the args you may need in <train.sh>, then run sh train.sh in terminal.

First change the args you may need in <test_fold_3D.sh>, then run sh test_fold_3D.sh in terminal.

Currently, we support CUDA version.

You can use one gpu or more than two gpus, just change CUDA_VISIBLE_DEVICES and nproc_per_node.

If our paper or code inspire your researches, please cite us by

@article{wang2023med,
  title={Med-tuning: Exploring parameter-efficient transfer learning for medical volumetric segmentation},
  author={Wang, Wenxuan and Shen, Jiachen and Chen, Chen and Jiao, Jianbo and Zhang, Yan and Song, Shanshan and Li, Jiangyun},
  journal={arXiv preprint arXiv:2304.10880},
  year={2023}
}

Great thanks for the code from Swin-Unet, timm.