For the diagnosis of skin lesions, the risks of artificial intelligence fail in the skin of color


In image analysis to detect potential pathologies, artificial intelligence (AI) shows enormous promise in multiple fields of medicine. But technology in dermatology is likely to fail in skin color if training does not specifically treat these skin types, according to Adewole S. Adamson, MD, who outlined this topic at the American Academy of Dermatology Virtual Meeting Experience.

Dr. Adewole Adamson

“Machine learning algorithms are only as good as the resources through which they learn. Without representation of skin-colored individuals, we run the risk of creating a new source of racial disparity in patient care.” Adamson, he told the meeting, assistant professor in the dermatology division of the Department of Internal Medicine at the University of Texas at Austin.

Diagnostic algorithms that use AI are typically based on deep learning, a subset of machine learning that depends on artificial neural networks. In the case of image processing, neural networks can “learn” to recognize objects, faces, or, in the realm of health, disease, from exposure to various images.

There are many other variables that affect the accuracy of deep learning for diagnostic algorithms, including the depth of the layer through which the process distills multiple information inputs, but the number of inputs is critical. In the case of skin lesions, machines cannot learn to recognize features of different skin types without exposure.

“There are studies that show that dermatologists can excel at detecting skin cancers by AI, so this will be an increasingly powerful tool,” Adamson said. The problem is that “there has been very little representation in darker skin types” in the algorithms developed so far.

The risk is that AI will exacerbate an existing problem. Skin cancer in darker skin is less common, but is no longer diagnosed, regardless of AI. The rate of 100,000 men in the United States melanoma it is approximately 30 times greater in white men than in black men (33.0 vs. 1.0). Among women, the racial difference is smaller, but still huge (20.2 vs. 1.2 per 100,000 women), according to U.S. data.

Because of the under-representation of darker skin in studies conducted so far with AI, “one of the arguments is that skin cancer is not something important in darker skin types,” Adamson said.

It can be the other way around. The relative frequency with which skin cancer occurs in the black population of the United States could explain a low level of suspicion and ultimately delay diagnosis, which in turn leads to worse outcomes. According to one analysis extracted from surveillance, epidemiology and the final result (I WILL SEE) database (1998-2011), the proportion of patients with regionally advanced or distant disease was almost double (11.6% vs. 6.0%; Pg <0.05) in black patients, relative to white patients.

Not surprisingly, given the importance of early diagnosis of cancer in general and skin cancer in particular, the average survival of malignant melanoma in black patients it was almost 4 years lower than in white patients (10.8 vs. 14.6 years; Pg <.001) for nodular melanoma, the same study was found.

In humans, bias is reasonably attributed in many cases to judgments made about a small sample size. The AI ​​problem is analogous. Adamson, who does published research on the potential of machine learning to contribute to healthcare disparities in dermatology, cited work done by Joy Buolamwini, a graduate researcher in the media lab at the Massachusetts Institute of Technology. In one to study realized, the failure rate of AI facial recognition was 1% in white men, 7% in white women, 12% in black-skinned men, and 35% in white men. women of colored skin. Adamson said fewer contributions of colored skin are less likely to occur.

The potential for AI-derived racial bias in the diagnosis of the disease increases and becomes more complex when contributions beyond the image are included, such as past medical history. Adamson warned of the potential for “bias to be introduced” when social, cultural, or other differences that differentiate one group of patients from another are not taken into account. However, for skin cancer or other diseases based on images only, he said there are solutions.

“We’re in the early days, and there’s time to change that,” Adamson said, referring to the under-representation of colored skin in AI training sets. In addition to including more skin types to train recognition, creating AI algorithms specifically for dark skin is another potential approach.

However, his key point was the importance of recognizing the need for solutions.

“AI is the future, but we need to apply the same rigor to AI as to other medical interventions to ensure that technology is not applied in a biased way,” he said.

Dr. Susan Swetter

Susan M. Swetter, MD, professor of dermatology and director of the pigmented lesions and melanoma program at Stanford University Medical Center (California) and the Cancer Institute, agreed. As a person who has tracked the progress of AI in skin cancer diagnosis, Swetter recognizes the potential of this technology to increase diagnostic efficiency and accuracy, but also called for specific studies for skin color.

The algorithms “have not yet been adequately evaluated in people of color, especially black patients in whom the dermoscopic criteria for benign versus malignant melanocytic neoplasms differ from those with lighter skin types,” Swetter said in a interview.

She sees the same solution as the one proposed by Adamson.

“Efforts are needed to include colored skin in artificial intelligence algorithms for validation and additional training to avoid possible overdiagnosis or underdiagnosis damage in patients with darker skin,” he noted.

Adamson does not report possible conflicts of interest relevant to this topic. Swetter had no relevant disclosures.

This article originally appeared on, which is part of the Medscape professional network.

Source link