Dataset Card for Conditional Polyp Diffusion

Dataset Summary

The Conditional Polyp Diffusion dataset provides synthetic gastrointestinal (GI) polyp images along with segmentation masks, generated using a two-stage diffusion modeling framework. The dataset is aimed at mitigating the challenges of data scarcity and privacy in medical imaging, especially for supervised polyp segmentation tasks.

Stage 1: Improved diffusion model generates synthetic segmentation masks.
Stage 2: Latent diffusion model generates corresponding realistic polyp images, conditioned on the masks.

This dataset enables training and benchmarking of polyp segmentation models, improving generalizability and reducing dependence on scarce annotated real data.

Supported Tasks and Leaderboards

Image-to-Image Translation: Generating realistic medical images from segmentation masks.
Semantic Segmentation: Supervised training of segmentation models for polyp detection.

Languages

The metadata and documentation are in English.

Dataset Structure

Each sample includes:

A synthetic GI polyp image.
A corresponding segmentation mask.

The images are generated to mimic the distribution of Kvasir-SEG masks and HyperKvasir polyp appearances.

Data Splits

The dataset contains:

1,000 synthetic masks
1,000 corresponding synthetic polyp images

Dataset Creation

Curation Rationale

Due to privacy and annotation constraints in medical imaging, the dataset addresses:

Lack of large-scale annotated datasets for polyp segmentation.
Need for diverse, high-fidelity training data for robust CAD systems.

Source Data

The improved diffusion model is trained on the Kvasir-SEG dataset’s segmentation masks. The conditional polyp generator is trained using these generated masks to create realistic polyp images.

Annotations

Masks are generated via diffusion models conditioned on prior distributions.
No manual annotations are provided; instead, generated masks are verified for similarity and diversity.

Usage

The dataset is intended for research in:

Medical image generation
Semi-supervised and supervised segmentation
Evaluation of synthetic data utility in clinical tasks

Evaluation

Three segmentation models (UNet++, FPN, DeepLabv3+) were trained with various combinations of real and synthetic data. Results demonstrated that using synthetic data can improve model performance, particularly with DeepLabv3+ achieving a micro-imagewise IoU of 0.7751.

Citation

@inproceedings{machacek2023mask,
  title={Mask-conditioned latent diffusion for generating gastrointestinal polyp images},
  author={Macháček, Roman and Mozaffari, Leila and Sepasdar, Zahra and Parasa, Sravanthi and Halvorsen, Pål and Riegler, Michael A and Thambawita, Vajira},
  booktitle={Proceedings of the 4th Workshop on Intelligent Cross-Data Analysis and Retrieval (ICDAR '23)},
  year={2023},
  doi={10.1145/3592571.3592978}
}

License

Apache License 2.0

Datasets:

deepsynthbody
/

conditional-polyp-diffusion