DynaFormer: Dynamic Dual-Attention Transformer For Medical Image Segmentation

Updates

• December 1, 2025: Initial release with arXiv.

How to use

The script train.py contains all the necessary steps for training the network. A list and dataloader for the Synapse dataset are also included. To load a network, use the --module argument when running the train script (--module <directory>.<module_name>.<class_name>, e.g. --module networks.DynaFormer.DynaFormer)

Training and Testing

1. Download the Synapse dataset from here.

2. Run the following code to install the Requirements.

   pip install -r requirements.txt

3. Run the below code to train the DAEFormer on the synapse dataset.

   python train.py --root_path ./data/Synapse/train_npz --test_path ./data/Synapse/test_vol_h5 --batch_size 20 --eval_interval 20 --max_epochs 400 --module networks.DynaFormer.DynaFormer

   --root_path [Train data path]

   --test_path [Test data path]

   --eval_interval [Evaluation epoch]

   --module [Module name, including path (can also train your own models)]

4. Run the below code to test the DynaFormer on the synapse dataset.

   python test.py --volume_path ./data/Synapse/ --output_dir './model_out'

   --volume_path [Root dir of the test data]

   --output_dir [Directory of your learned weights]