Evidence that the most sophisticated artificial intelligence (AI) to date can automate the diagnosis of some cutaneous lesions shows that skin cancer screening can potentially be transformed — even beyond the clinical setting, according to experts.
Deep-learning algorithms could increase the number of screened patients "and prioritize limited resources for patients with the highest risk for cancer," write Roger S. Ho, MD, from the Ronald O. Perelman Department of Dermatology at New York University School of Medicine, New York City, and colleagues in a column published online August 22 in JAMA Dermatology.
"Moreover, their use would likely increase the number of dermatology referrals and streamline patient visits," the editorialists say, pointing to recent reports that deep-learning algorithms approach dermatologist-level classifications of cutaneous tumors (Nature. 2015;521:436-444; J Investigat Dermatol. 2018;138:1529-1538).
An advanced subset of machine learning, deep-learning algorithms use multiple neural network layers to develop human capabilities for problem-solving and learning. Such algorithms are already showing evidence of outperforming humans on specific tasks, such as image recognition — used in self-driving cars, in drones, and now in the diagnosis of some cancers.
However, say Ho and colleagues, many questions remain about the use of deep learning to diagnose and manage disease. They note that without the leadership of dermatologists, "the tremendous potential of deep learning to change the field may never be fully achieved."
Dermatologists must get actively involved in discussing where deep learning "fits into the skin cancer screening paradigm," the editorialists suggest. Defining malignancy thresholds could be a good way to alert patients to see a dermatologist, for example.
"Future discussions may also center on the content patients are provided if given a diagnosis, or whether an app should disclose diagnoses to patients at all," they say.
When approached for comment, Adam Friedman, MD, professor of dermatology and director of Supportive Oncodermatology at George Washington School of Medicine and Health Sciences in Washington, DC, agreed that many questions and quality assurance issues must still be resolved. "These applications do not have a place in clinical practice as of yet," he told Medscape Medical News.
However, Friedman predicted that AI will improve early detection of both nonmelanoma and melanoma skin cancers, and that when combined with human clinical expertise, it could create "a super healthcare organism."
"I do feel like this will ultimately be a great extension to our clinical prowess, a cyborg-esque improvement that enables us to make faster management choices with potentially less morbidity," explained Friedman, a fellow of the American Academy of Dermatology.
However, he added, "my fear is that the public will use this technology in place of seeing a dermatologist."
As previously reported by Medscape Medical News in January 2017, a new deep-learning algorithm, aided by advances in computer science and large datasets, was as capable as clinicians for classifying skin lesions as malignant or benign. In October 2017, researchers cautioned that using technology to target skin cancer mortality required a whole-body surveillance approach because human-driven applications cannot identify which single mole or lesion to target.
AI is also being explored in other clinical field. In November 2017, researchers reported that the deep-learning algorithm model CheXNet could detect pneumonia from chest radiographs better than practicing radiologists could. The 121-layer neural network (by comparison, machine learning has a single-layer neural network) output the probability of pneumonia as well as a heatmap to identify the clinically suspicious areas. When the algorithm was expanded, it once again proved more accurate than clinicians at diagnosing 13 additional pathologies, such as cardiomegaly and pleural thickening.
Vision of the Future
As the algorithms become better and better, is there a chance that they could replace clinicians to some extent?
That is a vision of the future suggested by William Harless, MD, PhD, a medical oncologist and hematologist, in Sydney, Nova Scotia, Canada, who is founder and CEO of Encyt Technologies Inc, a cancer research center. He wrote in the comments section on a recent Medscape Medical News article about the IBM Watson algorithm for cancer treatment.
Physicians will ultimately be replaced with healthcare "experts" with a 2-year vocational degree, he suggested. It is these professionals, not doctors, who will implement the decisions made by the computers, he says. To add insult to injury, Harless adds that under such as system, "patient care will improve."
"As much as one might want to believe otherwise, algorithm-generated decision-making will ultimately favor computers over human beings given the ability of the computer to analyze far more data than the human mind is capable of analyzing," said Harless. "AI technology is just too broad and too sophisticated for a different outcome to be predicted. I would not bet the other way."
Ho and colleagues and Friedman have disclosed no relevant financial relationships. Harless is founder and CEO of Encyt Technologies Inc.
JAMA Dermatol. Published online August 20, 2018. Abstract
Follow Medscape Oncology on Twitter: @MedscapeOnc
Medscape Medical News © 2018 WebMD, LLC
Send comments and news tips to firstname.lastname@example.org.
Cite this: AI Set to Transform Skin Cancer Screening - Medscape - Aug 29, 2018.