OCTDiff: Bridged Diffusion Model for Portable OCT Super-Resolution and Enhancement

https://neurips.cc/virtual/2025/loc/san-diego/poster/117693

Requirements

conda env create -f environment.yml

Training Script

For running the training / inference script, see template

Dataset Curation

The train/val/test set should by default in this format:

/Dataset_Root
├── /Train
│   ├── /LowRes
│   └── /HiRes
├── /Val
│   ├── /LowRes
│   └── /HiRes
└── /Test
    ├── /LowRes
    └── /HiRes

The data_splitter is useful to curate such path structure. Please consider modifying the customized dataloader otherwise. To implement loss function with weights, a .csv file is needed.

Model Training

To switch on / off the ANA module, parse False here:

    params:
      ana_on: True

Multiscale Cross Attention is inherently integrated, based on CrossFusion and x-transformer.

Acknowledgements

Our code is based on BBDM and OpenAI Guided Diffusion, Improved Diffusion.

Citation

@inproceedings{tian2025octdiff,
  title={OCTDiff: Bridged Diffusion Model for Portable OCT Super-Resolution and Enhancement},
  author={Tian, Ye and McCarthy, Angela and Gomide, Gabriel and Liddle, Nancy and Golebka, Jedrzej and Chen, Royce and Liebmann, Jeff and Thakoor, Kaveri A},
  booktitle={The Thirty-ninth Annual Conference on Neural Information Processing Systems},
  year={2025}
}