banner of LMP department's DP&AI initiative

Adeyi envisions “amazing technologies” advancing into the pathologist’s arena

|
William Hoffman
7 minutes
Patients will benefit as pathology enters a new era

Digital pathology and Artificial Intelligence image“It’s only going to get better.”

That declarative statement gets bandied about a lot these days, often in the context of artificial intelligence, AI.  For the purposes of this story, it belongs to anatomic pathologist, liver disease expert, and former LMP professor Oyedele Adeyi now at the University of Alabama at Birmingham.

Oyedele Adeyi photo

Last fall, Adeyi gave an LMP Grand Rounds entitled "Next generation pathology in healthcare delivery:  Where we are and what's next."  Given Adeyi’s national and international reputation in the field, his perspective on things including the history of pathology, its development from ancient times, its theories and tools through the ages, and its contemporary practice and challenges carry special weight.  But is the field responding with sufficient dispatch to forces at work that require medical disciplines to adapt in ways that roil the familiar way of doing things? Pathology has incorporated molecular biology and genomics into its research and clinical practice.  Can it do the same with digital pathology but on a much broader scale?

Adeyi made his “it’s only going to get better” remark in the context of one of the machine demonstrations he’d recently seen by a digital technology company that uses AI to simulate standard histopathology staining.  The company’s virtual staining technology is currently under review by the U.S. Food and Drug Administration (FDA).

“I was born a skeptic, so I’m skeptical about everything,” Adeyi said in an interview.  “I saw the demo of virtual staining at an exhibit in Las Vegas by more than a few companies.  Things look better than they actually are.  That’s their job.  Having said that, most of the things we do today start by being too good to be true and people making claims that, yes, more often than not, fail to meet the sales pitch.  But then, as the concept is developed, it always gets better.  The demo I saw was quite impressive.”

What really caught Adeyi’s eye was the apparent ability of a trained algorithm, deployed on a virtually stained cell nucleus, to distinguish between two nuclear markers, the Ki-67 protein and the estrogen receptor (ER).  Ki-67 gene expression is found only in rapidly dividing cells in many solid tissue tumors. The Ki-67 proliferation index is used to measure how fast cancer cells in a tumor are dividing.  Abnormal ER signaling is found in hormone-related tumors in women.  Both Ki-67 and ER are common features of breast cancer and important treatment and outcome predictors.  Testing is typically done by immunohistochemical staining of biopsy tissue.

“I’m a complete novice here,” Adeyi said as he reflected on the demonstration he witnessed.  “I’m just trying to translate what my understanding is.”  Each protein molecule has its own fluorescent characteristics that it expresses in the nucleus.  But when the company trained the AI for Ki-67, for example, it didn’t just use the nuclear stain to extract Ki-67 features.  It also used mass spectrometry, he said.    The AI algorithm registers both the fluorescent characteristics revealed by nuclear stain and the molecular characteristics revealed by mass spec.  It is trained to see that these characteristics together match those of Ki-67 and not other nucleus-located breast cancer biomarkers such as ER.   “To me that makes a lot of sense.”

Combining AI nuclear staining with mass spec data is an example of what Adeyi calls biomarker multiplexing.  “That is the thing that is really coming in now,” he said. “You’ll be able to look at all these markers one by one, but on the same (unstained) tissue section.  Let’s say I’m looking for Ki-67.   Today I would have a tray of about ten slides each with a single marker.  So now, we are going to take one section, unstained, and be able to look at different markers individually or multiplexed, looking at which cells project multiple signals concurrently.  And then I can choose what color I want each protein to display and not be restricted to the predetermined detection method by IHC [immunohistochemistry], for example.”

Virtual stains comparable to histochemical and immunohistochemical stains (eg. hematoxylin and eosin or H&E, Masson’s trichrome, pan-Cytokeratin) and multiplex visualization of the tumor cell in its matrix context and immune cells in the vascular microenvironment “are amazing technologies,” Adeyi said, adding that he’ll probably be retired by the time they’re fully adopted in the clinical workflow “but they’re really exciting.  You’ll be looking at all those biomarkers spatially in a multiplex environment.”

Exciting and expensive.  Not a lot of pathology departments have the financial resources to keep up with what commercial instrument and software businesses and pathology foundation models built on proprietary datasets are rolling out.  “What I think is going to happen is what is already happening,” Adeyi said.  “Industry is going to partner with academic departments to develop this.  But the money, the funding, the market share, and the proprietary ownership, that’s going to belong mainly to industry.”

When new technology starts getting the attention of the FDA, “the only thing that would happen is that it’s going to get better,” Adeyi observed.  As an example, the FDA approved Paige Prostate, its first AI product for digital pathology images, in 2021.  In 2025, the agency granted Paige breakthrough device status for its Paige PanCancer Detect, an AI-assisted diagnostic application that enables pathologists to detect suspect elements and patterns associated with cancer across a wide range of tissues in multiple organs.  A prospective clinical trial of prostate cancer showed that Paige PanCancer Detect “significantly reduces IHC costs while maintaining diagnostic safety standards.”

Digital pathology, AI, and related technologies are going to change the way pathologists practice “whether we like it or not,” Adeyi said.  In his Grand Rounds talk, he paused his slide of a standard bifocal microscope to reflect: “This tool, as much as we love it, is no longer going to be sufficient for interacting in the new environment.  We can argue and say this is going to take a while but pretty much everything I showed you is getting to the point of FDA approval.  That means it’s already being used.  Maybe not by everybody, but it’s already being used.  So, this [digital pathology/AI] is going to be the tool for the near future.” 

Adeyi envisioned what he sees as the coming technological environment for pathologists.  “We are sitting in a pilot cockpit looking at a screen display, the laboratory information system.  Instead of a slide tray this is what you’re going to get.  Here is my case list.  I can transition from an H&E slide to immunohistochemistry on a single unstained section, for example, without ordering new stains, waiting a day or more for the new stains to return, reviewing, and maybe ordering a new set of stains.  I believe the time is going to come when we will be able to look at images that were obtained from spectral microscopy where we can select whether we want to look at HER-2 [a breast cancer biomarker] or ER / PR [estrogen receptor / progesterone receptor breast cancer biomarkers], where we can choose the stain that we want and the system is going to analyze the tissue and convert the cells that express the markers of interest without our having to make new stain requests [for immunohistochemistry or molecular tests].”

”Is it going to happen tomorrow?  Probably not.  But the technology is so robust to the point that it’s only going to get better and real,” Adeyi said.  “With the kind of technology available now, it’s no longer science fiction.  The time is coming when the tissue is going to go right from the body into the machine.  You generate images and do virtual staining and be able to obtain results in a short while. The workspace, the tools, standards, and the demands are changing, and rapidly, too.”

Here are the stories in our DP & AI series: