This research proposes a PDE-based trainable refinement module for tubular-structure segmentation to enhance shape continuity.

• Presentation Slides
• JSAI 2025 Abstract (Japanese)

• PDE-based refinement improves mask continuity while maintaining thin shape
• Our approach is applied to various architectures for tubular-structure segmentation because of post-process

• DRIVE dataset

• Ubuntu 24.04.2 LTS
• NVIDIA TITAN RTX(Memory: 24GB) * 4

1. setup the environment.

$ pip install -r requirements.txt

2. Download DRIVE dataset to /dataset.

3. Run /dataset/data_process.py to fit dataset for training models.

$ cd dataset
$ python data_process.py -dp DATASET_PATH -dn DATASET_NAME

4. train and test model.

$ cd pde-shape-refiner
$ main_ddp.sh