AI supported dermatology in skin of colour
Automation meant to address access to care may compound inequalities if biases in source data are not addressed (1,500 words, 7 minutes)
A lack of skin of colour images in medical texts and image libraries not only impacts medical education directly, but may also be introducing bias into automated systems designed to speed and simplify the diagnosis of skin conditions. This was a topic discussed by Dr. Jenna Lester during a presentation at the 7th annual Skin Spectrum Summit on Nov. 6.
Dr. Lester is an assistant professor of dermatology at the University of California San Francisco, where she founded the Skin of Color Clinic.
In her talk, Dr. Lester noted that augmented intelligence or artificial intelligence (AI), including machine learning, is increasingly being used as predictive tools to help dermatologists make diagnoses.
However, the photosets used to train the algorithms for these tools can be biased, she said.
“If we have a body of photographs that contains biases—when certain skin diseases are shown on certain skin tones, or if there is overall just a lack of representation of skin of colour—this could bias augmented intelligence or machine learning algorithms in a way that humans are also biased,” said Dr. Lester. “In this way, it could perpetuate existing inequities.”
Training these automated systems on biased data could even widen disparities by making it easier for certain groups to be diagnosed, but even more difficult for others.
Dr. Lester said that while much of the work on AI is done by computer scientists, doctors can help by raising bias in images as an issue whenever possible and by trying to make available as many high-quality photos of a diverse range of skin diseases in diverse skin tones as possible.
Part of the problem is that research from Western or high-income countries can be biased toward lighter-skinned populations, she said.
“We need to rely on our colleagues in India, on the continent of Africa and other places where there is ready access to a diverse range of skin tones, where there is representation of different disease prevalence,” said Dr. Lester.
Bottom Line: There is a lack of images of skin of colour in medical texts and image libraries. Augmented intelligence is being increasingly used as a tool in medical diagnosis, but those tools are trained on datasets that do not include a diverse range of images. This has the potential to widen already existing health disparities.
From the literature on automated or computer-assisted dermatology diagnosis
Machine learning for medical imaging: Methodological failures and recommendations for the future
The authors of this review article discuss several systematic challenges slowing down the progress of developing systems for computer analysis of medical images. These challenges include limitations of the data, such as biases, and research incentives, such as optimizing for publication. In this paper, they review roadblocks to developing and assessing methods for developing and utilizing machine learning and note that there is a potential for biases to impact the process at many stages.
Also discussed are ongoing efforts to address these challenges and recommendations for further efforts.
Lack of transparency and potential bias in artificial intelligence data sets and algorithms: A scoping review
This scoping review included research articles published between Jan. 1, 2015, and Nov. 1, 2020. The review authors evaluated studies describing the development or testing of deep learning algorithms for triage, diagnosis, or monitoring of skin disease using clinical or dermoscopic images.
Among the 70 unique studies included, more than one million images were used to develop or test artificial intelligence (AI) algorithms. However, only roughly one-quarter of the images were publicly available. Only 14 of the studies included descriptions of patient ethnicity or race in at least one data set used, and only seven studies included any information about skin tone in at least one data set.
The authors observed that a majority (64.3%) of studies developing algorithms for identifying skin cancer met gold-standard criteria for disease labelling. They also found that public data sets were more often cited, which could mean that public data sets will contribute more to algorithm development and benchmarks.
Artificial intelligence to support telemedicine in Africa
The authors of this paper note that teledermatology has obvious advantages in African countries that have limited medical care, long geographical distances, and relatively well-developed telecommunication sectors. While various working groups have supported the use of AI-based technologies to assist local physicians, the authors write that ethnic variations represent a challenge in the development of automated algorithms. They conclude that improving the amount of available clinical data is necessary to further improve the accuracy of the AI systems. That goal can only be achieved with the active participation of local health care providers in Africa as well as the dermatology community, they write.
The continuous development of a complete and objective automatic grading system of facial signs from selfie pictures: Asian validation study and application to women of three ethnic origins, differently aged
In this study, researchers asked two questions. One, how well an AI-based grading system compared with dermatologic assessments of 16 Asian-specific signs of facial aging (seven previously evaluated and nine new signs), in a sample of 112 Korean women. Two, to confirm the validity of those signs in 1,140 women of three ancestries—African, Asian and Caucasian.
The authors found that the 16 signs significantly correlated with clinical evaluations made by two Korean dermatologists. When applied on a larger scale to women of different ethnicities, good accuracy and reproducibility of evaluations were seen, though these varied depending on ethnicity. The facial signs dealing with skin pigmentation were found to be more relevant among Asian women than African or Caucasian women. The automatic gradings were even found of slightly higher accuracy than the clinical gradings.
These sorts of automated, reproducible assessments, which correlate well with medical grading, could be of great value to clinical research including epidemiological studies, the authors conclude.
VIDEO: AI in Dermatology: Promises and Pitfalls | DASH Invited Lecture Series
At the intersection of skin and society
A new Canadian national database tracking the lived experiences and unique challenges of Black youth related to academia will be unveiled later this year, according to an article in New Canadian Media.
This database will be built upon preliminary findings of research that began last October at the University of British Columbia. Those findings will be presented at the Congress of the Humanities and Social Sciences, which runs from May 12 to May 20, 2022. They show that structural barriers and lack of culturally-adequate support programs are major factors alienating Black youth from the academic world, according to the article.
The lead researcher on the project, Anette Henry, is a Professor in the Department of Language and Literacy Education, University of British Columbia.
Prof. Henry is quoted explaining that for Black youth, alienation from Canada’s academic environment begins in high school.
“If a kid feels alienated from school anyway… the child might feel that university is actually not for them.”
“We talk about inclusivity,” she said, “but we haven’t figured out how we really need to be inclusive.”
Prof. Henry explained to the news outlet that throughout their lives, Black youth are confronted with messaging that “there’s no point really going to university to reach your goals.” This may take the form of ads minimizing the importance of higher education, the history of systemic barriers to Black people accessing higher education, or the lack of Black representation in academia.
The research initiative, which consists of virtual and in-person monthly mentoring meetings with groups of grade 10 and 12 students, aims to counter that messaging by contributing to a larger conversation on what inclusivity is.
All team members in the initiative—professors, researchers, psychiatrists, neurosurgeons, and undergraduates who mentor high school students—are Black. This choice helped build a sense of self-worth in the participating students, Prof. Henry said.
“That really resonated with them,” she said, “that there’s a sense of possibility.”
The national database that is being built based on this data will allow researchers to access Canada-specific information that until now has come mostly from the United States, according to the article. Prof. Henry said being forced to rely on American data is problematic, as it does not necessarily reflect the situation in Canada.
This week
May 15 to 21 is National Emergency Medical Services Week in the U.S.
May 21 is World Day for Cultural Diversity for Dialogue and Development
Something to think about in the week ahead…
In two weeks
This coming Monday is Victoria Day in Canada, and Skin Spectrum Weekly will be taking the week off. The following Monday, May 30, we return with coverage of a talk by Dr. Bolu Ogunyemi on cultural competency in dermatology and how medical education can support that. Dr. Ogunyemi is Assistant Dean of Social Accountability and a Clinical Assistant Professor of Medicine in the Faculty of Medicine at Memorial University of Newfoundland & Labrador.