---
license: cc-by-nc-4.0
language:
- en
tags:
- data-centric
- data-cleaning
- quality-assessement
size_categories:
- 100K<n<1M
task_categories:
- image-classification
pretty_name: CleanPatrick - Data Cleaning Benchmark
configs:
- config_name: Combined Annotations
  data_files:
  - split: test
    path: "combined_annotations.csv"
- config_name: Label Errors
  data_files:
  - split: aggregated
    path: "label_errors.csv"
  - split: metadata
    path: "label_errors_meta.csv"
- config_name: Near Duplicates
  data_files:
  - split: aggregated
    path: "near_duplicates.csv"
  - split: metadata
    path: "near_duplicates_meta.csv"
- config_name: Off-Topic Samples
  data_files:
  - split: aggregated
    path: "off_topic_samples.csv"
  - split: metadata
    path: "off_topic_samples_meta.csv"
---
# CleanPatrick: A Benchmark for Data Cleaning

Welcome to **CleanPatrick**, the first large-scale benchmark designed for data cleaning in the image domain. 
Built on the Fitzpatrick17k dermatology dataset, CleanPatrick is a dataset for measuring the performance in detecting three major data quality issues: 
**off-topic samples**, **near-duplicates**, and **label errors**.

![CleanPatrick Teaser](./assets/CleanPatrick_Teaser_compressed.png)

## Overview

CleanPatrick consists of dermatological images annotated with over **500,000 binary labels** across **three data quality issues**:

1. **Off-topic Samples**: Images that are irrelevant to the dataset, such as non-dermatological content or images with no visible skin diseases.
2. **Near-Duplicates**: Highly similar images that may be caused by transformations, resolutions, or multiple views of the same condition.
3. **Label Errors**: Images with incorrect labels, including mislabeling and rare conditions mistakenly classified.

This dataset provides a realistic test bed to benchmark data cleaning strategies for image datasets, particularly in the medical domain.

![Data Quality Issues](./assets/DataQualityIssues.png)

## Key Features

- **Real-World Contamination**: Unlike synthetic datasets with artificially induced errors, CleanPatrick contains naturally occurring issues that reflect true real-world contamination found in dermatology image datasets.
- **Expert Annotations**: The dataset was annotated by medical crowd workers with expertise, and results were validated by medical professionals to ensure high-quality ground truth.
- **Evaluation Framework**: Along with the dataset, CleanPatrick provides an evaluation framework for benchmarking methods to detect data quality issues, offering standardized metrics to compare various cleaning strategies.

![Benchmark](./assets/benchmark_results_barplot.png)

## Dataset Details

- **Total Number of Images**: 17,000 dermatology images
- **Annotation Volume**: 500,000 annotations from 933 medical crowd workers
- **Categories of Data Quality Issues**:
  - **Off-Topic**: 4% of the images
  - **Near-Duplicates**: 21% of the images
  - **Label Errors**: 22% of the images

The dataset is available as a set of image-label pairs, with each image labeled according to whether it suffers from one or more of the three data quality issues.

## Installation

To load the dataset using the HuggingFace `datasets` library:

```python
from datasets import load_dataset

dataset = load_dataset("Digital-Dermatology/CleanPatrick")
```

## Citation
If you use this dataset in your research, please cite the following paper:

```bib
@article{
  groeger2025cleanpatrick,
  title={CleanPatrick: A Benchmark for Data Cleaning in Medical Imaging},
  author={Gr\"oger, Fabian and Lionetti, Simone and Gottfrois, Philippe and Gonzalez-Jimenez, Alvaro
    and Amruthalingam, Ludovic and Goessinger, Elisabeth V. and Lindemann, Hanna and Bargiela, Marie
    and Hofbauer, Marie and Badri, Omar and Tschandl, Philipp and Koochek, Arash and Groh, Matthew
    and Navarini, Alexander A. and Pouly, Marc},
  year={2025},
}
```

## Acknowledgements
We thank the medical crowd workers and domain experts who contributed to this dataset. 
The creation of CleanPatrick was made possible by their efforts and the valuable annotations provided.
Additionally we want to thank Centaur Labs for their help in collecting large amounts of crowdsourced annotations.

## License
This dataset is released under the CC BY-NC-SA 3.0 license.