sfczaa/ddpm-derm-synthetic-gallery

Versioned synthetic dermatofibroma gallery

This repository publishes the exact, pre-generated gallery used by the educational HAM10000 portfolio demo. It is not medical data for diagnosis or treatment and must not be presented as clinically validated imagery.

Published version

The repository root must contain this directory unchanged:

epoch0100_seed0/
  _READY.json
  metadata.json
  synthetic_df.csv
  images/

For low-latency container builds, the repository also contains a deployment archive of that exact directory:

• File: gallery_epoch0100_seed0.zip
• Files inside: 503 (500 PNGs plus _READY.json, metadata.json, and synthetic_df.csv)
• SHA-256: da3d582082323728e2b0558c27e26af124c683dacf336915d1212acd8abd0bc5

The original directory remains the published source of truth. The archive is a deployment-only transport copy; it does not replace, curate, or modify the gallery.

The version contains 500 class-conditional dermatofibroma (df) PNG images. They were generated with the epoch-100 EMA checkpoint, seed 0, 50 DDIM steps, eta=0, and 64 × 64 output resolution. The manifest has 500 rows and no missing images. The web gallery follows manifest order; images were not hand-selected for display.

The demo validates the version directory, readiness marker, generation metadata, row count, filenames, and image existence. It does not run DDPM generation during requests.

Source, license, and attribution

These synthetic images are derived from a model trained only on the HAM10000 training split. HAM10000 Dataset © ViDIR Group, Department of Dermatology, Medical University of Vienna. The ISIC 2018 data page distributes the source dataset under CC BY-NC 4.0. This derived gallery is published under the same non-commercial license.

Reference: Tschandl, P., Rosendahl, C. & Kittler, H. The HAM10000 dataset, a large collection of multi-source dermatoscopic images of common pigmented skin lesions. Scientific Data 5, 180161 (2018). doi:10.1038/sdata.2018.161

Modifications include a fixed lesion-ID group split, 64 × 64 preprocessing, class-conditional DDPM training, EMA checkpoint selection, and deterministic DDIM sampling. No endorsement by the original dataset creators or ISIC is implied.

Limitations

Synthetic images may contain artifacts, reproduce training-data biases, or fail to reflect clinically meaningful diversity. In the formal matched-585 experiment, using this gallery's synthetic data did not improve downstream df classification over real-image oversampling. The result is suggestive only and is neither statistically significant nor clinically validated.