AI’s Next Frontier — Dermatology? How It May Help Close The Health Gap For Dark-Skinned Patients

AI technology has been part of our daily lives and routines long before ChatGPT became a household name. It’s used whenever you unlock your phone with your face. It tailors your Netflix and TikTok dashboards. It analyzes traffic and gives you an estimated time of arrival on Google Maps.

But would you trust a computer program to diagnose your skin problems?

That’s the big question facing the field of dermatology. According to a Pew Research Center survey last year, nearly two-thirds of U.S. adults say they would definitely or probably want AI to be used to help detect skin cancer in the doctor's office. It's higher among younger people — 72% under 30 say they would want that.

Engineers are building systems that promise to help doctors and patients identify suspicious growths that could morph into skin cancer, as well as skin conditions like acne and psoriasis.

Some of these algorithms are for smartphone apps. Users can snap a picture of a rash or a bump and ask Google’s DermAssist for a list of possible diagnoses.

But most of the proposed systems are medical devices, meant to be used in a doctor’s office. There’s only ever been one AI diagnostic device approved by the U.S. Food and Drug Administration, called Melafind. It used a pattern-recognition algorithm to help a dermatologist decide whether a suspicious mole needed a biopsy. The FDA approved it in 2011 — for use in a dermatologist’s office — and it was touted as the future of early-stage melanoma detection. But the machine failed to recognize different forms of skin cancer lesions, and in 2017 the company announced that it was discontinuing the product after selling just 90 units.

Researchers hope newer AI tools under study can better detect skin problems than in the past.

“This is where machine learning is beautiful,” says Nyalleng Moorosi, a senior researcher at Distributed AI Research Institute. “You have a really massive, complicated problem trying to diagnose a disease.”

She says an AI system could be a powerful diagnosis tool as long as it works in tandem with an experienced doctor.

AI can provide options "and then you have an actual trained human make the final decision for you," she says.

Moorosi says it’s more concerning when these algorithms are put in the hands of less experienced dermatologists or with primary care doctors who weren’t trained in the nuances of diagnosing a skin condition.

AI tools "can be good assistants,” Moorosi says. “But the professionals also have to be empowered with the knowledge that these things can get things very wrong.”

This is especially concerning for dark-skinned patients, who often feel overlooked or are misdiagnosed because of medical racism. Studies have shown dermatologists can fail to recognize skin conditions on Black and brown patients.

Many of these systems also are being built without expert input. A 2021 study found that of 51 AI and skin cancer studies that year, only 41% included any dermatologists as co-authors.

This raises a key question: What if the systems being built perpetuate the same biases?

A Homogenous Field

Dermatology, like many medical fields, has a long history of racial inequality.

Most dermatologists are white. Only about 5% of dermatology residents were Black, compared to about 60% of white residents, according to data from the Association of American Medical Colleges. During the same period between 2020 and 2023, 24% of dermatologists identified as Asian and 7.3% as Hispanic.

The textbooks these dermatologists are trained on also often lack pictures of patients of color. Researchers evaluated several medical textbooks in 2020 and found only around 4.5% of the books showed images of conditions on dark skin. Conditions that look pink on lighter skin can often look brown on darker skin, or be nearly undetectable in comparison.

This has resulted in some dermatologists admitting that they are not comfortable diagnosing Black and brown patients. About half of 125 dermatology residents said in a study last year that they were not satisfied with their ability to spot skin diseases on darker skin.

While it was a small study, dermatologists say it mirrors the issues they often see in the field.

Dr. Jenna C. Lester says it’s very common for her to meet Black patients who sought her out because they felt overlooked and dismissed by white doctors.

“I spend a huge portion of an already short visit unpacking what I would consider to be medical trauma,” says Lester, a dermatologist at UCSF Health who founded the Skin of Color clinic.

“That kind of trauma manifests in different ways, but it often impacts health outcomes as well as a patient’s likelihood to seek out care again.”

While AI is often proposed as a solution to this disparity, the issue is that many of these AI systems are being trained on textbooks that include very few people of color.

The Race to Innovate

Dr. Roxana Daneshjou saw this issue in her own work. She first became interested in how AI could improve dermatology when she was pursuing her doctoral studies at Stanford University School of Medicine.

“Dermatology is a very visual field and it seemed like there was a real opportunity for leveraging these new AI technologies,” she says.

She knew there was a shortage of dermatologists, both globally and nationally. With her computer training, Daneshjou felt like the answer for this problem was simple: build AI tools that could help primary care physicians and non-specialists diagnose skin diseases. This need felt especially pressing in low-income areas, where the closest dermatologist could be miles away and booked for months.

But Daneshjou soon realized there was a problem: the datasets that the machines were training on didn’t reflect the people the dermatologists would be serving.

“Most of the datasets that were publicly available only showed disease on light skin,” she says.

With a team of researchers, she reviewed three AI diagnostic systems: ModelDerm, DeepDerm and HAM1000. All of these systems performed well on their original datasets, which drew heavily from the International Skin Imaging Collaboration, currently the largest public archive of skin lesion images.

In dermatology, skin tones are determined using the Fitzpatrick scale, which classifies skin color into six categories that range from from I to VI, with Type I being very pale and most likely to burn in the sun, and Type VI being very dark. When Daneshjou and her team personally curated a diverse dataset of conditions on darker skin tones, these systems performed worse, according to their 2022 Science Advances study.

AI is “a clever idea,” says Dr. Cheryl Burgess, founder of the Center for Dermatology and Dermatologic Surgery in Washington, D.C., and a member of the Skin of Color Society, an organization that spreads awareness about the dermatology needs of patients of color.

If there was more data showing skin diversity, "I think AI could be wonderful,” says Burgess. “Unfortunately, we don't have those images … it's definitely lacking, so therefore AI is lacking.”

The Problem with AI Bias

Algorithms are good at spotting and learning from patterns, even if that pattern leads to a bad outcome. So bias occurs when a system picks up on disparities and prejudices, assumes those are a feature rather than a bug, and then replicates them.

It’s “not hard to imagine” why this happens, says Hammad Adam, a researcher at the Massachusetts Institute of Technology who’s studying the intersection of AI, health care and inequity.

“The U.S., and the world at large, has had decades, centuries, of biases along racial, ethnic, and gender lines,” he says. “So any data that pulls from the real world — especially historical data — is going to have those biases built in.”

Possible Solutions

With a lack of datasets showing skin conditions on darker skin, researchers have proposed various ways to fill in the gaps with AI.

“It was concerning to discover that two of the most widely used image repositories for training AI models have a limited amount of images depicting people with tan, brown, and Black skin tones,” says Eman Rezk, a doctoral student at McMaster University in Canada.

She and a team of researchers trained an AI system to create synthetic images by melding together a darker skin tone with an image of a condition on lighter skin. But there are limitations to this. The generated images might not be able to capture all the nuances of how a skin lesion might appear on a person of color.

“It's very unsettling to me to put in a piece of generated data and expect to get a diagnosis,” Lester says. “AI is not going to solve societal issues that we have not devoted time and energy and resources into solving ourselves.”

Other solutions include using AI to audit AI. Daneshjou worked with a team of researchers to create a framework which can assess whether a medical textbook is diverse enough to be used as training data.

Similarly, both Google and Sony have proposed more nuanced skin color systems that they say engineers should consider looking at while building their AI systems: Google’s Monk Skin Tone Scale examines light to dark, while Sony’s scale considers red and yellow undertones. These are meant to build off of the Fitzpatrick scale.

But the real answer to these issues, Lester says, is that the dermatology field needs to focus on just getting more images of dark-skinned patients.

“We need to think of more creative strategies for getting diverse photos for diverse data to train these algorithms,” she says.

Lester imagines that this could work through partnerships with medical schools. Another method could be to consider compensating patients for the use of their images, especially if that photo is going to be used to train an algorithm that is going to make an AI company a lot of money.

Moorosi says ultimately dermatologists need to get involved with testing these tools. She likens engineers to carpenters, saying that they should be directed by the doctors who will use these systems and not the other way around. The technology will only improve if it learns how to identify problems on darker skin.

"We're using the system so it can learn, and not so that it can diagnose. We're not at a point where the system can offer knowledge," says Moorosi.