Researchers Train an AI Platform to Detect Skin Cancer

articles

Researchers have trained an AI platform to detect skin cancer. Here’s how their work could simplify the screening process for everyone.

Medically reviewed by Susan Kerrigan, MD and Marianne Madsen

Humans are imperfect. We get tired, we get distracted. We make mistakes. Computers, on the other hand, can analyze reams of data for days. Unfortunately, a computer’s output relies upon humans’ input. When it comes to diagnosing skin cancer, programmers have been stymied because in photos of malignancies, people of color are inadequately represented. Changing that became a vital first step as researchers trained an AI platform to detect skin cancer. Here’s what happened and how their work could simplify the screening process for everyone.

Race and skin cancer

All cancers begin with a genetic mutation that divides rather than dies. Cells with damaged DNA can develop into a tumor and spread to healthy cells. The goal of screening is to find dangerous cancers before that happens. The challenge is that many tumors are benign rather than malignant. Even cancerous tumors can grow so slowly that they won’t represent a threat for years, even decades. When cancer that isn’t dangerous is treated aggressively, the patient suffers while scarce health care resources are wasted. Of course, waiting carries its own risks. That’s why improving screening can be as important as improving treatments.

Skin cancer is the most common form of the disease –– which makes sense considering that it’s our largest organ. If you’re an average adult, you’re toting around eight pounds of the stuff. Laid flat, it would cover an area some 22 square feet. Whether you’re a conservative dresser or an active nudist, you’re exposing some skin to the sun nearly every day. Ultraviolet light, whether from our closest star or a tanning bed, is the primary driver of skin cancer.

Every day in the U.S., almost 10,000 people are diagnosed with skin cancer. Roughly 20% of the population will be diagnosed with the disease at some point in their lives. The subcategories are named for the cancer’s location –– with the less dangerous basal cell carcinoma (BCC) and squamous cell carcinoma (SCC) occurring in the basal or the squamous cells. Both are located near the skin’s surface. However, squamous cell carcinoma carries a greater risk since it frequently forms in areas that aren’t exposed to the sun including the genitals. More than three million Americans get these forms of skin cancer every year. It is highly treatable –– so long as it is identified early.

The more deadly form, melanoma, happens when there is damage to the color-producing cells –– the melanocytes. Like squamous cell carcinoma, this cancer can also occur in places where people are usually protected from the sun. So while the stereotypical skin cancer patient is fair skinned, both cancers also affect people of color. Research shows it’s far more deadly for them.

Every year, nearly 200,000 people in the U.S. are diagnosed with melanoma. Although this is the most dangerous form of skin cancer, patients have a 99% five-year survival rate–if it doesn’t spread to the lymph nodes. In fact, among White people the number of melanoma deaths has been declining.That progress hasn’t occurred with people of color. That’s partly because they are more likely to be diagnosed when the disease has progressed. One out of four African-Americans with melanoma are diagnosed after it has spread to their lymph nodes; 16% are diagnosed when it has also spread to other organs. This is the main reason why their survival rates lag that of Whites. The challenge is that people of color are more likely to have skin cancer in areas generally untouched by the sun. That makes early screening even more important. Unfortunately, early research using artificial intelligence was hampered by the lack of images of skin cancers in their community.

Training AI

First coined in 1955 by Stanford professor John McCarthy, artificial Intelligence (AI) was described by him as ”the science and engineering of making intelligent machines.” Today that means programming computers and other systems that learn from their mistakes –– just like human beings. And just like humans, computers need good information in order to make good choices. Early on, both images and data used to train AI to spot skin color weren’t representative of the larger population. According to University of Oxford’s Dr. David Wen, while AI programs have great potential as a diagnostic tool, “… it’s important to know about the images and patients used to develop programs, as these influence which groups of people the programs will be most effective for in real-life settings. Research has shown that programs trained on images taken from people with lighter skin types only might not be as accurate for people with darker skin, and vice versa.”

In Wen’s research examining data sets containing tens of thousands of images. For those that listed countries of origin nearly 80% were from Europe, North America, and Oceania exclusively. As he pointed out in his paper, when AI receives images primarily of Asian skin, it doesn’t perform as well with White skin; when its main source of images are from White people, it does a poor job diagnosing African-American patients.

In order to leverage AI’s promise and correctly diagnose members of Hawaii’s multiethnic community, researchers at the University of Hawaii Cancer Center focused on images of pigmented skin with lesions previously diagnosed as melanoma or non-melanoma. Given the Aloha state’s climate and its residents’ propensity for the outdoors, it’s a place with a high number of skin cancer cases. Among the state’s large multiethnic community, many go undiagnosed. The AI program was fed some 50 images obtained from the International Skin Imaging Collaboration dataset. Without any information on whether or not the images were of melanoma or non-melanoma, AI was correct around half of the time. Additionally, a trio of dermatologists viewing the same images matched the AI’s accuracy. However, combining the doctor’s assessment with the AI’s conclusions resulted in 100% accuracy. Although studies are ongoing, researchers are confident their work will soon lead to an improved diagnostic tool.

Being able to diagnose patients based on a photograph of a problem area means more than reduced time, pain, and expense for those being screened. It also greatly expands a dermatologist’s reach. A patient could be in an underserved rural community while the doctor remains in the city. Because it’s so simple, it may also increase the number of people being screened for skin cancer. If it’s accurate, it could also reduce mortality –– especially for patients from communities with lower survival rates who may feel more comfortable sharing a photograph than being examined in a doctor’s office.

References