Bias mitigation benchmark that includes two datasets

Description

ISIC-2020 is the largest skin lesion dataset divided into two classes -- benign and malignant. It contains 33126 dermoscopic images from over 2000 patients. The diagnoses were confirmed either by histopathology, expert agreement or longitudinal follow-up. The dataset was gathered by The International Skin Imaging Collaboration (ISIC) from several medical facilities. The dataset was used in SIIM-ISIC Melanoma Classification Challenge. In the images, the lesion is usually in the centre and well-visible. The examples of artifacts in this dataset that may introduce bias into the model include hair, frames, rulers, pen marks, or gel drops. Past research showed that frames are correlated with the malignant class and ruler marks with the benign \cite{mikolajczyk_biasing_2022}.

melanoma external malignant 256 (kaggle.com)

Gender classification dataset consists of cropped images of male and female faces. The data were collected from various Internet sources, most of which were extracted from the IMDB dataset. It contains 58658 images, with a similar distribution into female and male subsets. The authors of this article have discovered that glasses are a possible bias source, as actors wore them more often than actresses.

Gender Classification Dataset (kaggle.com)

Additionally, the presented dataset consists of masks that can be used for targeted data augmentation according to the method presented in the paper https://doi.org/10.48550/arXiv.2308.11386.

The research on bias reported in this publication was supported by Polish National Science Centre (Grant Preludium No: UMO-2019/35/N/ST6/04052). The authors wish to express their thanks for the support.

Dataset file

TDA-datasets.zip

1.3 GB, S3 ETag 697d36b88a6cfc2bd0e59dbb2170790a-3, downloads: 39

The file hash is calculated from the formula
hexmd5(md5(part1)+md5(part2)+...)-{parts_count} where a single part of the file is 512 MB in size.

Example script for calculation:
https://github.com/antespi/s3md5

download

File details

License:: open in new tab

CC 0

Public Domain Dedication
File embargo:: 2024-04-01

Details

Year of publication:

2024

Verification date:

2024-08-06

Creation date:

2022

Dataset language:

English

Fields of science:

automation, electronics, electrical engineering and space technologies (Engineering and Technology)
information and communication technology (Engineering and Technology)

DOI:

10.34808/d7pe-r837

Verified by:

Gdańsk University of Technology

Keywords

Cite as

Authors

seen 111 times

Search