AI Capabilities, Concerns 

Clifford Perlis, MD, discusses artificial intelligence in dermatology, including current capabilities and concerns as well as future opportunities. 

Clifford Perlis, MD, co-founder and former Chief Medical Officer of Belle.ai, Board-Certified Dermatologist, Keystone Dermatology Partners, King of Prussia, Pa.

“For so many years, we’ve talked about [AI] like it’s coming. It’s just down the road. It’s not just down the road. It’s here,” said Clifford Perlis, MD, who presented “Technology in Skincare – A Quiet Revolution” and moderated the panel “AI and Digital Therapeutics in Aesthetics” at the Science of Skincare Symposium in Austin, Texas.

“There’s a Nature publication from 2017 by a group at Stanford that showed that artificial intelligence was better than 20 of 21 dermatologists at diagnosing melanoma from a ceroscopy image.”

In the six years since, AI technology has continued to develop, said Dr. Perlis, which supports his first point: 

“AI is not something that’s coming. AI is already here.” 

According to Dr. Perlis, and his second point, one of the benefits of AI is that it can take on some of the time consuming and mundane tasks that dermatologists would probably rather not do.

“I don’t think of artificial intelligence as much as a threat so much as it’s going to be good at things that are boring or repetitive that dermatologists are not interested in doing.”

One example is quantifying response rates in clinical trials, said Dr. Perlis.

“When you’re developing a new therapeutic, like for acne or psoriasis or alopecia or vitiligo, the scoring [systems] are super tedious. I’ve yet to meet a single dermatologist who enjoys counting papules or comedones for a clinical trial. Yet that’s kind of a necessary evil. We need to quantify the response.”

AI is the perfect tool for this kind of work, he said. 

“The computer doesn’t get bored. And not only that, but… [it is] totally objective, totally uniform, totally repeatable. And so that’s one thing that artificial intelligence machine learning can do better than us. I think a lot of us are happy to hand it off.”

Computers don’t mind staying up all night to read papers, said Dr. Perlis. 

“And so just being able to sift through the tremendous volume of new publications and kind of filter it through and figure out what’s relevant, what’s important, take-home points. Again, it’s something that computers are really good at.” 

Dr. Perlis’s third Point: AI is getting better at the things dermatologists do. 

In a retrospective review, 2 researchers reported that an AI-based tool helped nurse practitioners and primary care physicians more accurately make dermatologic diagnoses, said Dr. Perlis. 

“It showed that it would have decreased the number of patients referred out for biopsies or consultations requested.”

Researchers also reported that using the AI tool added a few seconds to the evaluation.

Point number 4: The role of AI is broader than just dermatologic image analysis, said Dr. Perlis. 

“In the greater world, AI is doing a million different things… from self-driving cars to predicting the weather, and in dermatology there are lots of applications beyond just automated image analysis.” 

That includes customized patient information, decision support tools, and designing clinical trials and skincare products, he said.

“I don’t think it’s an exaggeration to say that we’re going to see a real paradigm shift with dermatology and skincare as artificial intelligence applications take hold.”

But while many tools have been and are continuing to be developed, most are not FDA cleared for clinical application, said Dr. Perlis. 

Importantly, AI algorithms need to have peer reviewed, published validation studies to show they really work and account for issues like bias, said Dr. Perlis. 

“The algorithms really can be only as good as the datasets they’re trained on and so many of our datasets come from individuals with lighter skin types. And so if we’re going to use them broadly, again, on the whole population, we need to include images of people with all skin types and different presentations. So that’s a huge issue.” 

Notably, it’s a recognized issue that AI developers are aware of and taking steps to address it, he said. 

Other issues include explainability and hallucinations, both of which developers are also working on. 

“Often times artificial intelligence is criticized for being a black box. You ask it a question, it goes into this black box, and an answer comes out and it’s not clear, certainly, how it comes to that answer, which makes it difficult for us to evaluate and say, ‘Well, does that answer make sense? Is it good? Is it bad?’” 

To explain how AI quantifies disease severity, for example, it can show outlines of disease involvement or a heat map, said Dr. Perlis. 

“In the area of automated image analysis, another approach to explainability is to say, ‘Well, I think this is tinea versicolor,’ and then show three pictures of tinea versicolor and say, ‘This is why the computer thinks it’s tinea versicolor because it looks like these three pictures.’” 

AI hallucinations occur when AI presents false or illogical information as facts.

“I think this is a problem that’s been identified and identifying it makes it easier to rectify it. A couple of strategies to mitigate this are using verified reliable sources for training material, so that you know you’re not just pulling random stuff, and then also keeping a human in the loop…who can say, ‘You know what, that’s not tinea versicolor. That’s a picture of a lion.’”

References:

  1. Esteva A, Kuprel B, Novoa RA, Ko J, Swetter SM, Blau HM, Thrun S. Dermatologist-level classification of skin cancer with deep neural networks. Nature. 2017 Feb 2;542(7639):115-118. doi: 10.1038/nature21056. Epub 2017 Jan 25. Erratum in: Nature. 2017 Jun 28;546(7660):686. PMID: 28117445; PMCID: PMC8382232.
  2. Jain A, Way D, Gupta V, et al. Development and Assessment of an Artificial Intelligence-Based Tool for Skin Condition Diagnosis by Primary Care Physicians and Nurse Practitioners in Teledermatology Practices. JAMA Netw Open. 2021 Apr 1;4(4):e217249. doi: 10.1001/jamanetworkopen.2021.7249. PMID: 33909055; PMCID: PMC8082316.