If melanoma has the reputation of being the most aggressively lethal skin cancer, cutaneous squamous cell carcinoma (cSCC) has traditionally been considered its less dangerous cousin. The cancer, which develops from the squamous cells in the outer layer of the skin, is typically less aggressive and less likely to metastasize.

But some cSCCs do go bad. And even though melanoma has a much higher death rate than cSCC, cSCC cases are so much more common – with over 1 million diagnoses per year in the U.S., about ten times the melanoma incidence – that they cause nearly as many deaths as melanomas do.

“Cutaneous squamous cell carcinoma is considered a cancer that rarely hurts anyone,” said Dr. Aaron Mangold, a dermatologist at Mayo Clinic. “But because so many people get diagnosed, ‘rarely’ starts to turn into a lot of people having complicated disease, needing health care and even dying. The issue is that it can be tricky to determine which cases could turn deadly.”

About 250,000 cases each year are considered “at risk,” according to Mangold. To help better predict the behavior of cSCC, he turned to Ravi Iyer, the George and Ann Fisher Distinguished Professor of Engineering, who is a leading artificial intelligence and machine learning expert in The Grainger College of Engineering. Along with their research teams, including Anirudh Choudhary, who is Iyer’s graduate student and a Mayo Fellow, Mangold and Iyer developed technology now being patented for a gene panel and the first comprehensive AI tool that evaluates risk to help pathologists predict outcomes in cSCC patients. 

The collaboration demonstrates the ability of AI to transform health care, according to Iyer.

“The tool can evaluate, with good success, whether the cancer will metastasize and even provide clues about when,” said Iyer, a professor of electrical and computer engineering who is also affiliated with the Coordinated Science Laboratory. “AI’s ability to help predict health care is central to so many things – better therapeutics, better care plans and better patient outcomes.”

Today’s pathologists are unable to establish connections between images of tissue captured by a microscope and the genetic changes of a tumor. The research team’s first step was to develop a gene panel that captures the critical genes driving aggressive cancer behavior, including disease spread and recurrence. Their current model predicts outcomes in at-risk patients with higher accuracy than any previous solutions. 

The Illinois/Mayo team integrated the genetic data within the AI tool to create a survival analysis model for outcome prediction. Specifically, the team used “weakly supervised learning” – a type of machine learning in which the training data have imperfect, incomplete or noisy labels – as the basis for the tool. They trained the model on de-identified images supplied by Mayo. 

One challenge with weakly supervised AI methods is that they often struggle with differences in image quality, such as variations in staining or visual noise. The researchers addressed this problem by training their model to learn from whole-slide diagnoses – looking at the overall tissue image – rather than depending on detailed labels for every tiny region of the slide. This approach helps the system focus on meaningful patterns while reducing the impact of image imperfections in small areas. The model also uses a ranking system that gives more weight to the most severe tumor regions, reflecting how pathologists prioritize the highest-grade areas when making a diagnosis.

“Our goal is to build something that doesn’t just give pathologists tissue information; it gives it in context,” Mangold said. “This tool could help direct the future of care, because it could help us understand for whom we should add a therapy or order a test, and when. We have a really big gray zone of people who are the most at risk of dying if they get this disease, and we currently don’t have much to offer them.” 

While cSCC develops on the outer layer of a person’s skin, other squamous cell carcinomas arise in other types of tissue, such as head and neck tissue or the lungs. The research team trained their AI tool across three types of squamous cell carcinomas; the tool was then able to outperform existing multiple-instance learning (MIL) methods (a type of weakly supervised learning methods) in making predictions for all three types, achieving a 3–9% improvement in grade classification and 5–20% higher accuracy on the higher-grade tumors that are the most challenging for pathologists. The approach also demonstrated a 16% higher sensitivity in localizing diagnostically significant tumor regions, compared to existing methods.

This collaboration was conducted under the oversight of the Mayo Clinic & Illinois Alliance, formed in 2010 to revolutionize patient care and improve human health through research, education and translation. Iyer said that Mayo is an ideal partner for this project, in part because the clinic has transitioned to entirely digital records over the past decade. It can therefore provide rich banks of information on which to train AI tools, including longitudinal data, such as imaging scans taken of the same tissue over a lengthy period of time. 

“Longitudinal information is relatively new to pathology, but it’s central to so many things,” Iyer said. “As we are able to see the behavior of tumors over time, we are better able to predict [their] future behavior. This could be the key to unlocking better therapeutics and outcomes.”

Mangold agrees. “Over 900,000 individuals in the United States have multiple cancers in their lifetime,” he said. “Currently we assess those tumors as independent events in the same individual. We are developing new models that allow us to assess multiple cancers over time in the same individual that will allow for more personalized diagnosis and management.”