banner of LMP department's DP&AI initiative

Turner foresees a bright future as the field goes digital

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

Digital pathology and artificial intelligence signature graphicDoes an image on a glass slide have more to tell than human eyes can discern while peering through a microscope?  If so, might that have implications for diagnosis of disease?  With their digital pathology project “Enhanced image analysis of colonic hyperplastic polyps with genetic abnormalities using artificial intelligence,” LMP associate professor Kevin Turner and his research team intend to find out.

No one knows why the rates of early-onset cancer — cancer diagnosed in those under 50 — are rising.   Colorectal cancer, the third most common cancer worldwide accounting for approximately 10 percent of all cancer cases and the second leading cause of cancer-related deaths, is a case in point.  Cases among adults younger than 55 doubled from 1 in 10 in 1995 to 1 in 5 in 2019, according to the American Cancer Society.  The disease tends to be more aggressive in young adults.  As a result, the U.S. Preventive Services Task Force in 2021 recommended that colonoscopy screening, which can be a life-saving procedure, begin at age 45 instead of 50. 

Kevin Turner photo

Turner has long experience over a microscope examining colorectal tissue including polyps, small cellular growths that form along the lining of the colon and are often extracted during colonoscopy.  What if molecular changes in polyps produce subtle alterations in the patterns and features of the tissue under analysis – alterations associated with possible progression to cancer -- that human eyes with the aid of “machine eyes” can detect and classify?

“In the colon there are many types of polyps,” Turner said in an interview.  “Some of them are cancerous.  Some are precancerous, which means they are neoplastic, which means they have a chance of turning into cancer if they are left alone.  And then there are non-neoplastic polyps such as hyperplastic polyps that have long been considered to be benign.  If you have the latter, your doctor will say, ‘See ya in 10 years.’” 

Recently, Turner said, researchers have begun to subtype hyperplastic polyps [HPs] based on histologic features.  They started to categorize these polyps by the way they appeared.  Then they ran molecular tests on them.  “Reportedly, some of these hyperplastic polyps harbor mutations in B-raf and K-ras genes, which are known to be precancerous mutations,” he said.  “So, a hyperplastic polyp with a B-raf mutation, that may not be a benign polyp.” Yet hyperplastic polyps, which are common and considered benign, are not screened for mutations.  Molecular screening of individual polyps is currently quite expensive.  That would be prohibitive in reflexively screening them, Turner said. 

“My idea is to see if we could train a model to screen out hyperplastic polyps that harbor B-raf or K-ras mutations by using image analysis,” Turner said.  “I suspect there are histologic features that will be seen more frequently in mutated HPs. Perhaps the cell nuclei are slightly enlarged.  Or maybe there are other changes that an algorithm can be ‘taught’ to pick that up. That would be a powerful tool because, as I said, hyperplastic polyps are so prevalent that it isn’t feasible to do molecular testing on all of them.” 

Turner imagines an artificial intelligence (AI) program that could be run on a hyperplastic polyp at a fraction of the cost of molecular testing and “may tell you, hey, this one has features of a mutated polyp.”  That finding could then be confirmed or ruled out with a molecular test.  “That’s a much more feasible workflow,” he said.  “The AI would be the screen and the molecular test would be confirmatory.  This could identify patients that may benefit from a shorter surveillance interval.” 

“The aim of this study is to identify a group of HPs [hyperplastic polyps] with genetic mutations and attempt to discern histologic features that are associated with mutations that an image analysis AI algorithm can be taught to identify,” Turner and his colleagues wrote in their grant application to the LMP divisions of Anatomic Pathology (AP) and Molecular Pathology and Genomics (MPG).   “Our end goal is to advance the understanding of clinically significant HPs and ultimately to develop a cost-conscious program that can screen them in lieu of expensive molecular testing.”

Turner said he himself has reached the age when colonoscopy is recommended.  “The screening age is 45 now.  I’m 45 and my doctor said I need one,” he said.  “If the worrisome trends we are seeing continue, I would not be surprised if that age changes to 40.  I’ve been practicing for 10 years.  In my first few years I rarely saw patients in their 30s or 40s with colon cancer. Now we are seeing it regularly. We don’t know why that is.”

Digital Pathology:  A big disrupter

Turner’s is a leading voice in the department for moving expeditiously into digital pathology with automated whole slide scanning, image management systems, machine learning- and artificial intelligence-enhanced specimen analysis software, workflow reengineering, and education.

“This is likely the biggest disrupter that we’ll see in pathology in a very long time,” Turner said.  “Since the beginning of pathology, the way we view histology has been the same.   Microscopes have gotten better.  Slides have gotten better.  Processes have gotten better.  But we still more or less shine a light at a slide and look at it.” 

The field of pathology looks to radiology to get an idea of where things are headed.  “You can kind of compare it to what happened in radiology years ago,” Turner said.  “The same thing happened with them.  The old way was to hold a film up to a light box and read it.  That’s how they did it.  Eventually, they discovered the advantages of digitizing their images, leading them to have the same discussions pathology departments are having now:  it’s really expensive, we need new equipment, new workflows, etc.  There was resistance to it at first, but now you don’t see radiologists holding film up to light boxes anymore.”  Digital imaging and scanning constitute the standard of care in radiology, he said.  “It will be so in pathology as well.  Once you have digitized images, you can do many more things with them.  There are challenges but we will all get there eventually.”

When radiology began its transition to digital, it had a lot of buy-in from hospitals, Turner said.  “There was more political awareness.  There was more money put into it.  So, many departments just said ‘We’re doing this.’  With pathology, it’s different because it’s a speculative investment and departments are not seeing the same level of buy-in from the powers that be, so it’s difficult to argue an ROI [return on investment] and to justify all the metrics people use to make investment decisions.  It is challenging to justify an ROI to offset the costs of expensive equipment, software, and staff when digital pathology cannot currently be billed for.”

That accounts for why most pathology departments are moving slowly into digital even though it seems to be an inevitability. “There are many advantages of a digital image over a slide,” Turner said. “They [digital images] last forever; slides degrade.  They can be shared instantly.    Instead of physically sending slides to a colleague and waiting a week for an answer you can look at the digital image with them the same day.  Digital images are fantastic for AI research.  They’re fantastic for trainee education and tumor boards.  They bring doctors and different groups together.” 

Turner gave the example of attending a virtual GI [gastrointestinal] tumor board and showing his colleagues cases from his office remotely through his microscope via Zoom rather than at a conventional in-person tumor board meeting.  “Digital pathology allows clinicians and pathologists to be in the same room for tumor boards where they can discuss patient cases and build camaraderie.  I believe it’ll even get to the point where clinicians can show interested patients their images,” he said.

Hardware and software team up to transform pathology 

Digital pathology shares some of the challenges associated with the emerging field of spatial omics.  As noted last year in “Spatial omics makes its debut in pathology” about the department’s spatial omics initiative, one of the problems is that the commercial instrumentation needed to advance the field is evolving so fast. Investment is made in an instrument and, in short order, there’s a superior instrument introduced, sometimes a far superior instrument.  The same goes for software.

Last year LMP acquired a Hamamatsu Photonics NanoZoomer S360 high-throughput whole-slide imaging (WSI) digital slide scanner.   The machine can scan 360 slides in a single run of 4.5 hours.  “It’s the best and brightest right now,” Turner said.  “They will keep coming out with faster scanners. Scanners that can hold more slides. Scanners that do more things, of course. But our scanner will be good for years to come.  If we need something new, we’ll still be able to use this scanner.   You’re always going to need a workhorse scanner.” 

Moving perhaps even faster in development than automated slide scanning machines like the Hamamatsu S-360 is the software that brings to life local and global patterns and features in tissue architectures. Advanced software is key to applying machine learning (ML) and AI to pathology specimens—to discerning, say, patterns in “benign” polyps that may turn out to be clinically important.

Indica Labs’ HALO, an image management system for digital pathology and clinical workflow, is among the leaders.  HALO AI is described as a “trainable AI” enterprise digital pathology platform, a highly sophisticated tissue classification system that “puts the power of deep learning artificial intelligence into the hands of pathologists and researchers.”  Last year, the U.S. Food and Drug Administration (FDA) issued a 510(k) clearance to Indica Labs for HALO AP Dx.  The platform is used for primary diagnosis of surgical pathology slides prepared from formalin-fixed paraffin embedded (FFPE) tissue scanned by the Hamamatsu S360. 

“We have purchased three HALO AI software packages from Indica that come with pre-learned modules, things they’ve made that can do specific tasks,” Turner said.  “You can take those and build AI to target whatever you’re looking for.  They have modules that can fragment cell nuclei or that can count things or quantify the intensity of the staining, for example. This system gives you plug-and-play applications that you can use for AI research.” 

Introduced in 2011, Telexy Healthcare’s QuPath is an open-source bioimage analysis software for digital pathology and whole slide image analysis and comes with tumor identification and high-throughput biomarker evaluation tools.  The software allows users to develop and share new algorithms to analyze complex tissue images.  “If you’re coming up with a completely new idea, and you’re training a model from scratch, it can be hard to find real-world examples or find someone who’s done it before you,” Turner said.  “That’s where QuPath comes in.  It’s a free open-source AI training software.  People can share what they’ve made on that platform.”

At the very tip

Turner notes that pathology is “just at the very tip” of the transition to digital.  The FDA approved Paige Prostate, its first AI product for digital pathology images, in 2021.  In Europe, approved AI products are now available for prostate cancer, breast cancer, colorectal cancer, lymph node analysis, metastasis detection, and relapse risk.  As of July 2025, the FDA had approved only five ML/AI-enabled medical devices for pathology including Paige Prostate compared to 116 for cardiology, 54 for neurology, 19 for hematology, 11 for microbiology, and more than 700 for radiology. The FDA approved the first AI-enabled medical device for radiology in 1995, yet the national employment picture for radiologists remains robust.

“You can imagine that when these algorithms come out and people start writing papers on how effective they are, eventually they are going to work their way into guidelines,” Turner said.  “When they are in the guidelines, you will either adopt the technology or you’ll have to send things out.”   

There are many obstacles to overcome, straits to navigate and bottlenecks to pass through before pathology becomes digital and automated in the way radiology is today.  “The best way I’ve heard it put is that when you invest in digital pathology, you’re investing in the department by keeping it modern and relevant,” Turner said.  “That’s ultimately where your return on investment is.”

“Enhanced Image Analysis of Colonic Hyperplastic Polyps with Genetic Abnormalities Using Artificial Intelligence”

PI:  Kevin Turner, DO

  • Co-PI:  Nathan Pankratz, PhD
  • Co-PI:  Dina El-Rayes, MD, PGY4
  • Co-PI:  Khalid Amin, MD, PhD
  • Co-PI:  Alessio Giubellino, MD, PhD
  • Co-PI:  Andrew Nelson, MD, PhD
  • Co-PI:  Brian Bagley, PhD

  • Co-PI:  Brook Gagner

Here are the stories in our DP & AI series: