Futuristic portrait merging human and digital elements, showcasing advanced facial recognition technology and AI integration in a striking blue hue over a woman's face

Skin in the Game

An IU researcher’s AI tool shows promise in predicting whether melanoma will return.

Matthew Harris Dec 17, 2024

IMAGINE YOU’RE ANXIOUSLY waiting in a dermatologist’s office. A week earlier you came to an appointment hoping to have a bump on your nose checked out. Now, you’ve returned for a follow-up — one where a clinician might tell you that mole is early-stage melanoma.

Often, treatment is straightforward: an outpatient procedure to remove the thin and superficial tumor. For 94% of people diagnosed, it will be their only brush with the fifth-most common cancer in the U.S.

But what if your provider could also tell you the probability that your skin cancer will return? What if the forecast was based on 150 variables undetectable to the human eye? And what if the results came back the same day?

At IU School of Medicine, Ahmed Alomari, MD, and his team are working doggedly to develop and train a machine-learning algorithm capable of doing just that.

The pace at which medical technology evolves isn’t slowing, but some tasks remain rooted in tradition.

“What we want is to develop an online tool,” Alomari said. “Anyone could upload a slide, and it would tell them the predicted 5-year and 10-year survival time for their patient. We hope that the tool will help them with risk stratification.”

The tools are advanced, but pathologists like Alomari still evaluate stained slivers of tissue under a microscope. His years of experience inform his analysis to render a diagnosis.

Yet even the keenest eyes are imperfect. They cannot detect every shade in the color spectrum or gauge the tiniest spatial relationships. And even if they could, our brains can’t quantify the relationship between those features and clinical histories — correlations required to build predictive models about whether melanoma will come back.

Now, machine learning, which uses AI to imitate the way humans learn, offers that avenue. With it comes the possibility that pathologists and dermatologists will have a fine-grained way to determine, for example, which melanoma patients need closer monitoring.

“Could there be something else that a machine and a high-capacity computer can look at, analyze and calculate in a short period to give us that extra information?” Alomari mused.

"What we want is to develop an online tool. Anyone could upload a slide, and it would tell them the predicted 5-year and 10-year survival time for their patient. We hope that the tool will help them with risk stratification.”

Ahmed Alomari, MD

So, Alomari, an associate professor of clinical pathology and dermatology, set out with his lab to develop a tool to handle that task.

They started by using scanners to magnify curated pathology slides 400 times to create extremely high-resolution images that show all the cells in a sample. Those scanners simultaneously culled and fed data on pathological features into a supercomputer. “These are features where you can’t look at it in a matter of seconds and draw a conclusion,” Alomari said.

Analyzing the images means breaking them into thousands of small squares, and the algorithm evaluates the shape of cells in that quadrant, the coloration of certain features like the nucleus, and how many cells are packed into the space. Then, it slowly zooms out, quantifying spatial relationships from surrounding grids.

Next, the algorithm calculates a score based on those physical features and clinical outcomes. Statistically speaking, a higher score indicates a greater likelihood of a relationship. Once that training was done, Alomari’s team asked its algorithm to evaluate another set of images. It accurately gauged the outcome in 90% of cases.

Unlike some approaches, Alomari’s team can also outline how its machine-learning algorithm reached conclusions. “It’s also not a complete black box,” Alomari said. “You can express what feature the machine is looking at and why it’s associated with certain risks or certain biological behavior of melanoma.”

Some of the model’s conclusions weren’t all that surprising, either. The thickness of a melanoma tumor still matters. Yet the process also detected the kind of nuance that a human might miss. For example, the more diversity there is in the size of cells, the better the outcome tends to be for the patient.

“That might sound bad, but it means there are lots of immune cells mixed in with tumor cells,” Alomari said. “They haven’t been crowded out by cancer.”

In the near term, Alomari is focused on validating the model’s performance in Stage 1 and Stage 2 patients. Part of it is need. Those patients still make up most new cases. In patients with advanced skin cancer, teasing out relationships between physical features and clinical outcomes can be trickier. A patient’s treatment regimen, for example, may trip up the model.

But in time — and hopefully with backing from an NIH grant — Alomari thinks his team could find solutions. “That way we could predict your response and survival based on the type of additional treatment you got to lower the chances melanoma recurs.”

Donor support can scientists pursue bold research that improves care for melanoma. Learn more and make a gift today.

Glossary Terms

Deep Learning: An offshoot of machine learning that utilizes artificial neural networks inspired by the structure and function of the human brain.

Large Language Model: A type of AI model that uses deep learning and is trained on huge sets of text data to recognize, understand and generate human-like text or other types of complex data. The model also be trained or fine-tuned to fulfill a task. ChatGPT’s ability to answer questions is an example.

Tags: Artificial Intelligence, Cancer, Research Findings, Pathology