Porcine Wound Forecasting Processed Data

This repository contains the processed data artifacts used by the Wound Forecasting project for longitudinal porcine wound-image generation and extrapolation.

The repository includes image-space data and the model-ready representations used by the DyneODE and River experiments.

Contents

.
├── manifest.json
├── images_256x256/
│   └── 256x256.zip
├── dyneode_inversions/
│   ├── inversions-00000.tar
│   └── manifest.jsonl
└── river_h5/
    ├── davinci_train.h5
    └── davinci_val.h5

Processed images

images_256x256/256x256.zip contains the processed longitudinal wound images used by the project.

The images are 256 × 256 pixels. Some historical experiment paths used the directory name 512x512; that name does not describe the released image dimensions.

DyneODE inversions

dyneode_inversions/inversions-00000.tar contains 1,650 serialized PyTorch latent tensors used by the DyneODE experiments.

dyneode_inversions/manifest.jsonl maps each source-relative latent path to its member name within the TAR archive.

The wound-domain StyleGAN generator required to decode these representations is released separately:

• https://huggingface.co/bridenmj/wound-stylegan

The final DyneODE checkpoint is available at:

• https://huggingface.co/bridenmj/wound-dyneode

River HDF5 data

river_h5/davinci_train.h5 and river_h5/davinci_val.h5 contain the model-ready HDF5 data used by the final River workflow.

The final River-only model weights and configuration are released at:

• https://huggingface.co/bridenmj/wound-river

River uses the external VQ-MUSE autoencoder. Its upstream weights are not duplicated in this dataset repository:

• https://huggingface.co/Emma02/vqvae_ckpts

Project code

Source code, configurations, evaluation utilities, and loading documentation are available at:

• https://github.com/bridenmj/wound-forecasting

Source dataset

These artifacts were derived from the public longitudinal porcine wound-healing dataset:

• Dataset publication: https://www.nature.com/articles/s41597-025-05921-w
• Dryad source: https://doi.org/10.5061/dryad.0rxwdbsbr
• Source dedication: CC0 1.0 Universal

Citation

If you use these processed data or derived representations, please cite the original dataset publication:

@article{zlobina2025high,
  title={A high-resolution temporal transcriptomic and imaging dataset of porcine wound healing},
  author={Zlobina, Ksenia and Yang, Hsin-ya and Kesapragada, Manasa and Lu, Fan and Gallegos, Anthony and Villa-Martinez, Guillermo and Alhamo, Moyasar A and Zhu, Kan and Recendez, Cynthia and Collins, Craig and others},
  journal={Scientific Data},
  volume={12},
  number={1},
  pages={1635},
  year={2025},
  publisher={Nature Publishing Group UK London}
}

Please also cite the associated Wound Forecasting paper when its publication information becomes available.

Data organization

The source images are organized longitudinally by pig, wound, day, and within-day burst. Consult manifest.json for the released dataset organization and dyneode_inversions/manifest.jsonl for TAR member lookup.

Intended use

The dataset is intended for research on longitudinal visual forecasting, generative modeling, representation learning, and wound healing progression.

Release status

This repository remains private while public-release permissions and licensing are finalized.