AI May Underperform in Spotting Cancer for Darker Skin


Source: Geralt/Pixabay

Artificial intelligence (AI) machine learning is emerging as an important assistive tool in cancer detection and diagnosis. The accuracy of deep learning not only depends on the algorithm itself, but also the quantity, diversity and quality of the dataset used to train the algorithm. A new study published in The Lancet Digital Health has found that darker skin is underrepresented in the publicly available image datasets used to train AI algorithms to spot skin cancer.

“This is the first systematic review to characterize publicly available skin image datasets, highlighting limited applicability to real-life clinical settings and restricted population representation, precluding generalizability,” wrote the researchers affiliated with University of Oxford, Churchill Hospital, University of Birmingham, Birmingham Health Partners, Health Data Research UK, National Institute for Health Research Biomedical Research Centre for Ophthalmology, UCL Institute of Ophthalmology, Databiology, and the Royal Berkshire Hospital.

For training AI deep neural networks used for computer vision, publicly available datasets are often used. In this study, the researchers analyzed 21 open access datasets that contained over 106,000 skin lesion images, 17 open access atlases, eight regulated access datasets, and three regulated access atlases, with the goal of performing a systematic review of publicly available skin image datasets used to train AI algorithms to identify skin cancer.

The researchers discovered that only 21 percent of the 14 datasets that identified the country of origin came from a region other than North America, Europe, and Oceania.

“There was limited and variable reporting of characteristics and metadata among datasets, with substantial under-representation of darker skin types,” the researchers reported.

There is a lack of transparency in the metadata reporting for characteristics such as ethnicity and Fitzpatrick skin type according to the researchers for not only in datasets for dermatology, but also for ophthalmology and radiology as well. This lack of metadata reporting may impact the generalizability of the AI machine learning algorithm. AI deep neural networks trained on certain types of image data may underperform on underrepresented images in the training dataset.

“Quality standards for characteristics and metadata reporting for skin image datasets are needed,” concluded the researchers.

Copyright © 2021 Cami Rosso All rights reserved.

Steve Liem

Learn More →