AI Answers a Call for Help in Cancer Detection

“Signs of a deadly tumor might not be noticeable until the very last specimen on a biopsy,” Mossel says. “Fatigue can have serious consequences, too. But if an AI algorithm can do the heavy sorting work, then a pathologist is left to examine the troubling slides with a clear head.”

In 2017, the World Health Organization (WHO) magnified a major and global opportunity in healthcare through new guidance emphasizing the benefits of early detection and diagnosis of cancer. The 40-page report concluded that programs that “facilitate timely diagnosis and improve access to treatment” would both save money and improve outcomes across countries of every income stratum.

“Diagnosing cancer in late stages, and the inability to provide treatment, condemns many people to unnecessary suffering and early death,” noted Dr. Etienne Krug, a director at the WHO at the time. Implementing WHO’s guidance, he continued, would be “less expensive to treat and cure cancer patients and… result in more people surviving.”

While WHO guidance included elements of more robust public education and healthcare financing, much of the burden of earlier diagnosis would fall on pathologists. A primary care physician or an oncologist might order an initial tissue biopsy, but it is a pathologist working behind the scenes who is ultimately responsible for confirming the presence of cancer through visual examination of that tissue. The challenge: there aren’t enough pathologists in the world. Workloads are increasing for those in practice, labs are understaffed, and medical schools aren’t producing enough new grads to replace those retiring from the field.

“Pathologists have been facing this Sisyphean task of increasing demand for their services and yet, each year fewer professionals are working in the field. As a result, there has been a building push to digitalize pathology,” said Joseph Mossel, CEO and co-founder of Ibex Medical Analytics, the leader in AI-powered cancer diagnostics.

“A decade earlier, artificial intelligence lacked the power to triage, much less accelerate a pathologist’s workflow. But toward the late 2010s, my co-founder, Dr. Chaim Lihart, and I saw the computing bottleneck starting to give. Processing speeds and machine learning technology were hitting their hockey stick growth, and the ability to dissect and understand images was improving rapidly. We took that opportunity to build a company that used AI to make sure that patients gain access to an accurate and timely cancer diagnosis from the get-go.”

A Hard Problem for Discerning Eyes to Solve
The field of pathology can trace its origins back to the tenth or eleventh century when doctors used lenses to get a better sense of what was happening in tissue. But it wasn’t until the mid-1800s and the invention of the microscope that pathologists gained a true picture of the tiny structures that make our bodies tick.

Yet when Mossel and Linhart founded Ibex, pathologists still relied on this technology—glass in tubes—to identify abnormalities in a cell’s nucleus, mitochondria, and accompanying structures. Unlike textbook diagrams, which render these components with clear boundaries and colors, pathologists are sorting through far murkier images. A pathologist must remain vigilant through hours of scanning and inspecting structures surrounding the cells, noticing subtle conditions like calcium build-up or poor vascularization, sorting through clumps of cells bunched to one side, or sometimes counting hundreds of cells in an image to categorize a particular disease state.

“Signs of a deadly tumor might not be noticeable until the very last specimen on a biopsy,” Mossel says. “Fatigue can have serious consequences, too. But if an AI algorithm can do the heavy sorting work, then a pathologist is left to examine the troubling slides with a clear head.”

For AI to produce accurate answers, the model needs a baseline of what an object, concept, word, image—anything—means or is. Humans develop the “ground truth” for an apple by repeatedly seeing and handling an apple; a machine arrives at a ground truth after being shown and told an image contains an apple multiple times and then building associations with the shape, color, and surrounding conditions of the apple and apple scenarios.

To train Ibex’s AI platform, named Galen—the company hired human pathologists, to “re-diagnose” thousands of histology slides from patients diagnosed with cancer. Like teachers who teach young kids to distinguish between the different letters and then build words and sentences out of them, the pathologists would annotate dozens of different cell types and then supervise Galen’s learning, testing how the algorithm performs when it is deployed on a new data set, until it hits the right accuracy levels.

Today, Galen can flag cancer cells across mammary, prostate, and gastric biopsies, distinguish between the cancer subtypes and inform its grading as well as recognize over a hundred malignant and non-malignant morphological features within these tissues. Whether for a machine or human, just tracking all the variations within this small sliver of the human body makes for an incredible task.

AI-assisted Reviews
Galen is not another physician; like most AI systems, it does not return results that proclaim, “This is an X” and “This is a Y.” Rather, it is a never-tiring digital assistant that helps the overworked pathology teams to triage and prioritize cases, detect cancer and other important findings, and then be the final arbiter of what’s happening.

To make the pathologist’s workflow more efficient, Galen generates a kind of “heat map” on the slide, coloring areas to show the pathologist which cells and structures it found the most concerning. AI models often deliver results as “final” with no clear rationale on how they arrived at the answer, whereas Galen offers what all good physicians and scientists crave: explainability, with the algorithm pinpointing the different morphologies one by one and sorting how distinctive they are, mimicking how a pathologist is trained to look at tissue.

“We’re moving a step further upstream in the workflow, and in addition to supporting the pathologists on their clinical work, our technology opens the door for new efficiency gains that optimize laboratory work ” Mossel says. “It’s been a breath of relief for our customers. Galen’s allowing pathology teams to meet the growing demand for cancer testing.”

At the company’s start, Galen offered “second reads”—or, in other words, reviewing a pathologist’s work and flagging possible errors such as missed cancers or tumors that were not graded correctly. Now, Galen is also doing “first reads,” where it reviews pathology slides before the pathologist and flags the areas the physician should examine more closely, helping them to report cases faster and more accurately.

“We’re moving a step further upstream in the workflow, and in addition to supporting the pathologists on their clinical work, our technology opens the door for new efficiency gains that optimize laboratory work ” Mossel says. “It’s been a breath of relief for our customers. Galen’s allowing pathology teams to meet the growing demand for cancer testing.”

Making an Impact
In 2021, Ibex’s AI technology gained “Breakthrough Device Designation” by the U.S. FDA and it is now the most widely deployed AI solution in pathology, serving health systems, laboratory networks and hospitals around the world. Since then, one of the areas Ibex has delivered the most impact has been in improving the diagnostic quality and reducing error rates at smaller pathology labs and community hospitals, which may not have a full complement of specialized pathologists on staff.

“Cancer research moves forward in quantum leaps and keeping up with the latest findings across all tissue types is too huge a challenge for generalist pathologists. They simply can’t know it all,” says Mossel. “Our technology serves as an equalizer, providing the utmost accuracy to all patients, regardless of where they live or what is their healthcare plan. A great example is a nationwide deployment of our AI solution in Wales, which demonstrated a pronounced improvement in the quality of cancer diagnosis and care across the country.”

“Cancer research moves forward in quantum leaps and keeping up with the latest findings across all tissue types is too huge a challenge for generalist pathologists. They simply can’t know it all,” says Mossel.

With more than $100 million in funding from investors like Dell Technologies Capital, the company continues to scale to meet the growing need for innovative diagnostic solutions, inking partnerships with Roche and AstraZeneca, as well as integrating into the clinical workflow at major hospital systems like the University of Pittsburgh Medical Center and Hartford Healthcare. A recent Nature paper lauded Galen‘s accuracy in detecting breast cancer, and in March, Galen became one of the winners of the UK’s NHS Artificial Intelligence in Health and Care Awards.

This traction is a small step towards Ibex’s larger goal to transform cancer diagnosis and become the de facto standard for AI in pathology.

“When we started, physicians questioned the effectiveness of AI, but there’s little doubt anymore,” said Mossel. “We reached a point where pathologists who experience our solution are no longer willing to move back to diagnosing slides on their own. Working with AI has become part of their everyday routine.”

Even with the significant progress they’ve delivered in cancer detection, the Ibex team isn’t yet content. The company plans to expand its product portfolio to include detection for additional tissue types and cancer biomarkers. They aren’t going it totally alone, though, and are partnering with some of the biggest leaders in the industry including Roche and Philips. The hope is that by creating end-to-end AI-powered clinical workflows, catching more cancers earlier will help to achieve WHO’s goals of better cancer care and improved patient outcomes across the board.

“When we started, physicians questioned the effectiveness of AI, but there’s little doubt anymore,” said Mossel. “We reached a point where pathologists who experience our solution are no longer willing to move back to diagnosing slides on their own.”