February 2025—Digital pathology, artificial intelligence, and anatomic pathology computer systems—seven participants in a Dec. 10, 2024 online roundtable talked with CAP TODAY publisher Bob McGonnagle about their experiences, plans, and predictions. Large academic center practices and small pathology practices—they considered all perspectives. Here is what they told us.
CAP TODAY’s guide to anatomic pathology computer systems begins here.
Coleman Stavish, tell us why you think we asked Proscia, the provider of the Concentriq platform, to be part of this roundtable on anatomic pathology computer systems.
Coleman Stavish, cofounder and chief technology officer, Proscia: We’re seeing that digital pathology has moved beyond the talking and planning phase and into the operational phase. The scanner and software technology has matured considerably over the past couple of years. Our Concentriq platform sits at the center of bringing these solutions together. Some large labs in the U.S. and around the world have gone almost 100 percent digital, if not fully digital. So there’s a blueprint now for the market to see how it works at scale, and Proscia can speak to this.
This is the first time we’ve dealt in depth with digital pathology and AI for anatomic pathology systems. It’s becoming a reality in many places and in part reflects the excitement around AI in pathology to simplify the complexity of diagnosis, but it also reflects the shortage of surgical pathologists in the United States. Dr. Rashidi, can you comment on that and what it means to you at UPMC?
Hooman Rashidi, MD, MS, associate dean of AI in medicine, University of Pittsburgh School of Medicine; professor and endowed chair of experimental pathology research, and executive vice chair of computational pathology, University of Pittsburgh Medical Center; executive director, Computational Pathology and AI Center of Excellence, University of Pittsburgh: Because of the shortage of pathologists and staff, the leadership of hospitals and institutions is more receptive to investing in complementary frameworks, specifically digital pathology, that could be incorporated into workflows. It’s a cost to them now. But if they see it as incorporated with AI, where the return on investment is there, at least from a three-to-five-year strategic plan, an argument can be made that a digital pathology framework is needed because without it there is no AI. Once you set up digital pathology and a digital transformation, you can incorporate the AI, which then fuels the adoption and return on investment.
About 90 percent of our workforce is outside of academia and don’t have scanners, so the number one question is how to entice them to get on board knowing they have to learn a new innovation, and as with any innovation you’ll have a J curve. The difference now is you can argue that the dip down of the J curve can be flattened in digital pathology, at the least, with AI. End users will have an incentive to use digital pathology tools because their workflow speed and accuracy will be enhanced, especially with regulatory frameworks that show you may have decreased risk-management issues.
It was clear from the start that digitizing pathology was a means to set up a system that could have AI and other computational pathology. You couldn’t be in AI without the digitization. Dr. Sinard, talk about that and about there being more than one up and down in the history of digital pathology.
John Sinard, MD, PhD, professor of pathology and of ophthalmology and visual science; medical director, pathology informatics, Yale University School of Medicine: It’s important to keep in mind that the landscape looks different whether you’re in a large academic medical center or a smaller practice, and the majority of pathologists are in small practices. It affects one’s view.
With respect to the up and down of digital pathology, initially there was an excitement about the technology; some places were doing it even though at the time there was no good return on investment. For a long time, the whole slide scanner was a technology looking for a problem. It was the solution to something, but people tried to make up the problem it was supposed to solve. There were select use cases, remote locations that had limited pathology access. But that didn’t generalize, and I think it accounted for the second lull.
The recent rise in interest is linked to the potential of AI, but the potential is far from realized. The ROI is still a difficult argument. Those who are using improved efficiency and speed as justification are probably using AI in a way that extends beyond the regulatory approval, which at least for now is restricted to use in conjunction with traditional microscopy rather than as a replacement for pathologist examination of the slide.
I’m concerned that initial attempts to deploy AI into practice will increase review time rather than decrease it because the algorithms, wanting to err on the side of sensitivity versus specificity, are going to identify areas that pathologists otherwise would not have spent much time looking at but now will because AI identified them as suspicious.
There are unanswered questions and much will depend on the individual algorithms and how places elect to deploy them.
Sam Terese of Alverno Laboratories was probably the first lab executive in a large system to plunge into digital pathology. His argument is that the ROI will come later but will be there because we’re going to have a consolidation of pathologists under systems, either tightly or loosely. Joe Nollar, give us your reaction to what you’ve just heard.
Joe Nollar, associate vice president of product development, XiFin: John’s description of the evolution of adoption of digital pathology and AI resonates. In the initial phases, with ChromaVision Medical Systems and its scanners and then Aperio and BioImagene, which have since been purchased by Leica Biosystems and Ventana, respectively, it used to be niche, research-oriented or reference labs that were leveraging digital pathology in niche revenue models. That was exciting and fun and was an insight into what was to come. But widespread adoption took a great deal of time and still isn’t there.
AI has been the fuel that has ignited a burst of interest in digital pathology. We’re seeing it permeate community pathology practices at a much broader level. In our demos we’re routinely asked how we integrate digital pathology and digital pathology workflows into our lab information system. It is an AI-driven argument now, although they see the other benefits of digital pathology from a remote access standpoint, and it’s going to take off from here. AI algorithms will get better and more refined, create even more interest and necessity, and help alleviate the stress of the pathology shortage.
Suren Avunjian, is digital pathology an important component of what people want you to speak about as it pertains to your system?
Suren Avunjian, cofounder and CEO, LigoLab Information Systems: Absolutely. Pathology groups, mostly of eight or more pathologists but even smaller practices, are looking to adopt digital pathology. They’re focused on how to seamlessly integrate advanced viewers, scanners, and other technologies into their workflows. The value lies not just in having cutting-edge technology but in embedding it deeply into the pathology workflow to create a cohesive and efficient solution.
For this integration to reach its potential, the industry needs to address persistent challenges. Many viewer technologies still rely on older standards like HL7, which limit their ability to integrate seamlessly. To unlock the next phase of evolution, these systems must adopt more robust integration points, such as standardized APIs [application programming interfaces] and web services. These technologies enable deeper, more meaningful integration that can transform digital pathology from a standalone tool into a core part of an optimized workflow. At LigoLab, we’re driving this evolution by ensuring our platform facilitates these deeper connections.
Leigh Boje, what do you hear about digital pathology and AI from your customers and potential customers?
Leigh Boje, pathology product manager, Orchard Software: We’re being asked more questions during our demos. Which vendors do you integrate with? What does that look like? Our existing customers are researching affordable hardware, scanners, and image management systems. Many vendors are finding a more affordable price point for the laboratories we’ve historically serviced. When it comes to scanner throughput, you don’t need to buy a minivan when you need a sedan, so there’s a lot of shopping for different scanners. Image management systems that are agnostic to scanners have an advantage because it allows labs of the size we traditionally serve to piece together an affordable solution.
To speak to adoption, whoever starts building their data set has a leg up as the AI algorithms evolve. The data is already there to take advantage of the evolution in AI, so the faster these practices can start digitizing their current workload and go back into their archive, the better the position they will be in in the data economy.
Diana Richard, pathologists feared losing their job at the introduction of digital pathology, and AI also has raised anxiety. Has the anxiety shifted? Is it diminishing as there are more role models?
Diana Richard, associate vice president national accounts, XiFin: As the adoption rate improves, we’re going to see the anxiety come down. We’re trying to build consumer confidence in digital pathology and how we leverage AI to improve patient care, data storage, and long-term valuation and study of information and disease states beyond a single event.
There will be driving forces behind not only creating a sense of confidence but also countering industry pressures. We are understaffed in pathology in the United States, so there is a significant cost for recruitment, if you’re even able to recruit a pathologist into your region. Digital pathology helps solve for some of that.
We saw at a digital pathology conference in 2023 a positive and enlightening debate between the CMOs of PathGroup and of Cleveland Clinic. It provided an example of how the perception of digital pathology continues to be varied among laboratories. PathGroup has fully adopted and deployed digital pathology, and they use it in recruitment and with consults and real-time reads as standard operating procedure. Cleveland Clinic has approached it more conservatively because, at the time, from their perspective the ROI was not material enough to justify full deployment. Even in well-funded national institutionalized laboratories, the way they perceive the ROI, the risk of adopting digital pathology and subsequently AI, are on different sides of the scale.
Dr. Rashidi, fill us in on the vision you have at UPMC and what attracted you to join UPMC.
Dr. Rashidi (UPMC): The number one reason was my close friend, Dr. Liron Pantanowitz, became the chair of the Department of Pathology. I bring Liron up—and Dr. Matthew Hanna, whom we recruited recently to join our group—because we’ve come to realize that to make an impact we have to consolidate our silos. UPMC is a mega player in informatics and digital pathology and AI, but until a couple of years ago, it was silo based—different people doing different things. We want to make sure we have the right infrastructure, so we created a new division called computational pathology and informatics and set up a new AI center called CPACE, which is our Computational Pathology and AI Center of Excellence. We’re recruiting computer scientists and physician scientists who deal with various R&D elements within the AI center. Matthew is leading a new center called DPRC, the Digital Pathology Research Center, which is around the instrumentation and scanning elements as well as various AI tools.
Another large section within our division is our clinical lab, which is called the pathology image analysis laboratory. Whatever we build, either from CPACE or with our vendor partners, gets deployed in this CLIA-approved image analytics lab. Many of our AI tools, our own or from vendors, are running within the image analytics lab, and our pathologists use them not as a replacement for their decision but as a complement.
We’ve come to realize two major things. One is there are early wins we need to go after, which is the unimodal or less complicated AI tools as they’re incorporated into our digital pathology space. These are tools that tell someone whether or not they have prostate cancer. Or using digital pathology as a framework without the AI. But the near-future landscape is not going to be unimodal; it will be multimodal. So we’re building an infrastructure to support that so our whole slide images, or even non-whole slide images, along with radiology and laboratory medicine data, and other clinical and molecular data, can crosstalk and give a bigger picture so it becomes an integral part of not only our department but the institution as a whole as an invaluable resource that could be part of the DRG landscape.
Our model is also a hybrid model, meaning we feel the best solutions are hybrid solutions. That means when the ROI allows, you set up something with the right vendor partner. When the ROI doesn’t allow or the vendor can’t provide it, we do homebrew versions of the same thing, and that’s why we have our CPACE, DPRC, and other centers in place.
The infrastructure needs to be built carefully because not all instruments are alike, and the regulatory landscape doesn’t always understand that right now and I don’t think it will in the near future. As an example, you build a non-FDA-approved AI tool now and sell it to us for use within our digital pathology framework; it’s complementary to our diagnostic process and we may choose to use it. If the vendor decides to take it through an FDA process, which many of them do, then you’ve made that process static now, meaning that FDA has approved it around a particular scanner, for example. If my institution uses a different scanner, you have forced my arm to not want to use your tool because of potential risk-management issues. These internal elements should be thought about up front, or we should at least help our regulatory colleagues come up with a process that is more nimble and enables addendums, additions, or enhancements with greater ease when a new framework is approved. If I have an AI tool that’s FDA 510(k) approved and I’m using different instrumentation, there should be a process whereby the vendor can get additional validations around that 510(k) with additional instruments. It’s not an easy task. It’s a complex landscape because there are many players.
And we don’t know where we’ll be with laboratory-developed tests in the coming year.
Dr. Rashidi (UPMC): That’s right. Most people think LDTs relate just to laboratory-developed tests and not to AI, but we know many people are starting to think about certain AI tools as LDT-like elements. Certain SaMDs [software as a medical device] may fall under that. Because of all this, we’re also concentrating on the education piece. We need to have the right education to democratize AI literacy and its proper deployment and usage to as many people as possible to increase adoption.
Coleman Stavish, people from the outside might assume if you were to develop the kinds of products and systems you’ve developed at Proscia, it would be easy to deliver the same message to every pathology laboratory. But they’re all different. Tell us about your experience with that.
Coleman Stavish (Proscia): There’s not one single message that will resonate with all reference labs, academic centers, and hospital systems. In the immediate term, with the technology currently available, everyone should look at developing their own digital pathology framework, and that will look different if you’re a pathology practice, a large reference lab with many contributing laboratories, or a more centralized academic center or small community practice. What you might achieve with today’s digital pathology is going to look different, as is the value proposition. The LIS, scanner instrument, or software and AI technology you acquire might also look different depending on what you’re trying to achieve.
While the specifics will differ, an overarching message that is true for everyone is you can achieve near-term wins with digital pathology today, like being able to sign out from home, work remotely, have a more flexible schedule, cut down on logistics costs for the lab, and reduce turnaround time. You can achieve these in a six-to-12-month implementation project.
Leigh [of Orchard] made a good point about building a data set and setting yourself up for future success from an economic and precision medicine standpoint. Digital pathology is not just a different way of doing the same thing; we’re creating a new class of whole slide image data, whether it’s brightfield or multiplex fluorescence imaging, that provides one of the most detailed and direct profiles of disease. The data can be valuable for sponsored research with pharmaceutical companies and can be the foundation for developing AI that makes a pathologist more efficient at tasks they already do. But I also think this data will increasingly be used to build novel diagnostic and prognostic tests, adding new information to the precision medicine puzzle, and this is especially exciting.
Cernostics, later acquired by Castle Biosciences, has been doing this for a while. They developed a proprietary AI-based LDT called TissueCypher, which is able to provide a good prognostic score for esophageal cancer in patients with Barrett’s esophagus. That’s an example of an AI image-based precision medicine assay. It’s non–tissue destructive, it’s adding new information, and payers started paying for it. It’s not about replacing a pathologist’s read of an H&E. That still happens. TissueCypher provides an additive piece of information. This is just one example of how new and valuable data are being created, putting histology and pathology at the center of enabling a new approach to precision medicine.
John Sinard, would you say that within five years every pathologist will be practicing either in a system where they’re in a central hub of a large pathology enterprise or on a periphery, where we’re going to need tools to share cases at a subspecialization level and many of the needs and development will be driven by that distributed model? It’s not antidemocratic to think we’re going to have large centers like Pittsburgh and others helping community pathologists. That was part of the vision of many early adopters.
Dr. Sinard (Yale): Five years is optimistic and arguably aggressive. A lot of details need to be worked out. From the 10,000-foot view, all of this sounds great—we’re going to get all the experts to be able to deploy their expertise where it’s needed, and ideally you practice at the upper limit of your ability. When you start looking into the weeds, however, you’re crossing state lines, so there are licensing issues. Who’s going to pay for what? Is this out-of-state coverage? Are these Medicaid patients for whom anything out of state is not covered? Who pays the technical? Who pays the professional? How the details are worked out will be institution and regionally specific with respect to licensing. And we haven’t begun to speak to the regulatory aspects. How do you accredit such a distributed diagnostic system in a meaningful way?
I would say that by five years, everyone who is practicing pathology will fit somewhere in a network, and it may be loose or tight. We see consolidation and subspecialty pathology driving a lot of the currents in the field.
Dr. Sinard (Yale): That’s a fair statement. I have a somewhat biased view, being at a larger academic medical center. I don’t know to what extent that view would be shared by the two-to-three-pathologist practices in smaller settings. Those hospitals may not be targeted by large conglomerates wanting to absorb them into their networks.
Suren Avunjian, let’s say you’re in a group of five or six pathologists. You know you’re going to have to be absorbed, so you’re making choices around that. What are the top two or three tips every pathology group should bear in mind as they approach new software and hardware?
Suren Avunjian (LigoLab): The first step is to have clarity on your long-term vision. If you’re scaling from a group of five or six pathologists to 10, 12, or even 20, identify the functionalities your group will need to thrive in the future. Ask yourself whether the vendor you’re considering has the capacity to support this growth. Will its solution enable you to centralize your operations into a single source of truth, or will it leave you managing fragmented systems that could slow progress?
Modern software must not only meet operational needs but also help alleviate the pathologist and technologist shortages. Choosing technology that prioritizes automation can be a game-changer for operational efficiency.
Finally, consider where the new system places your group in terms of strategic capabilities and interoperability. Will it enable you to onboard new partners, integrate cutting-edge tools, and adopt technologies like digital pathology with ease? Evaluate the timeline for implementation—will this be a disruptive, yearlong project, or can it be accomplished with minimal downtime? The ability to implement swiftly and scale effectively is crucial to maintaining momentum during growth or transition periods.
Leigh Boje, what two or three things should people have top of mind as they look for a new AP vendor?
Leigh Boje (Orchard): Understand your current resource skill set. Many systems that have been sunsetting are on-premise, maintained with hardware, so evaluate if you want to continue to take responsibility for those costs and investments and if your vendor is able to continue to support that model. Can your vendor offload some of that cost by hosting in a cloud environment or train system administration services to maintain the software so you take some load off your IT folks? We offer the ability to do your own integrations so you don’t have to call us each time you bring in a new surgery center from which you receive specimens.
Also, make sure the system is able to capture and communicate the increasing amount of information that is needed to reach the higher reimbursement rates.
Joe Nollar, what two or three things should be top of mind as people look for new systems, regardless of their size, setting, or where they plan to go in a few years?
Joe Nollar (XiFin): Cloud-based solutions are becoming integral to laboratories. They offer scalability, flexibility, security, and cost efficiency. They don’t have to worry about yearly upgrades or maintenance of hardware and on-prem solutions. With all the security threats health care institutions have faced over the past years, security is something they’re trying to outsource to cloud-based vendors. They see the benefit of getting routine updates and upgrades, having application flexibility via configurations, and being able to accommodate multimodality testing across a broad spectrum of testing.
I appreciated Dr. Rashidi’s comment about how UPMC’s computational pathology department is focused on multimodality solutions. LIS vendors have to think about providing multimodality solutions that aggregate the discrete data not only for precision medicine reporting across modalities into comprehensive summary reports, but also to provide the discrete data for different computational pathology solutions.
That is our focus as labs are looking to advance into the next generation of lab information systems. It centers on not having to worry about the day-to-day maintenance of the lab system on-prem and on having an end-to-end solution that has a financial model fully integrated into the lab information system as part of providing comprehensive services and systems to laboratories.
Diana Richard, with the shortage of health care IT personnel, do you find yourself getting the last ticket at the bakery when you have a project, particularly if the system has a large, centralized IT group?
Diana Richard (XiFin): Our runway for our own implementations has pushed out, particularly in hospital settings where we work with IT departments that have multiple priorities and the lab is not at the top of the list. You’re going to have more challenges in bigger institutions around getting the budget and the authority and priority to implement the technology. There’s also red tape around the decision-making and approval processes.
Since February 2024, the security protocols have become exponentially more challenging to navigate, particularly with health systems or research institutions. Being able to work with them from a contracting perspective has become challenging. The point of bandwidth and IT is real but the same can be said about the many departments the lab will have to work with to get this type of approval in a large institutional system. We don’t face that challenge when we’re working with independent clients that have an IT consulting firm or in-house IT.
We’re starting to see the cost of digital pathology soften as suppliers get creative with how they offer solutions to customers. For example, if the vendors that clients are looking at purchasing digital pathology platforms from also offer instrumentation, reagents, or other solutions not necessarily related to digital pathology, when they renegotiate their contracts as a package, they’re able to get some costs of the digital pathology integration skimmed as a result of committing to purchase other items.
Coleman Stavish, what do you want to know about a system’s, lab’s, or pathology group’s IT infrastructure so you can prepare for a meaningful sales call?
Coleman Stavish (Proscia): We want to understand where they are and how they’re thinking about digital pathology. Are they thinking about going as fast as they can to 100 or near 100 percent digitalization? Are they looking at biting off smaller use cases like TC/PC splits or consultation first? Then we’ll look at if they have scanners. Do they have plans in the next 12 months to buy new scanners? If so, which ones? And what’s their LIS? Many times the answer about an LIS is, “We’re using this LIS but switching to this other LIS in 12 months, so what do we do about that?” We’ve worked with customers on creative implementation plans where they can get value today—maybe their slide barcodes have information that enables them to use digital pathology—and then we pencil in their bidirectional LIS integration with a digital pathology platform once they go live with their new LIS.
On the cybersecurity and IT infrastructure side, we want to know where they want their digital pathology images stored and software to run. While we continue to support and bring on new customers who deploy on-prem in an intelligent way, we’ve seen a big change in the past 18 to 24 months in how people think about cloud. There used to be hesitation, but we see more and more interest in a hosted cloud-based solution. We need to understand the pathology and IT departments’ goals and create a plan that speaks to both.
Dr. Sinard and Dr. Rashidi, there’s a need for pathology reporting that has some discrete data and a version of the report that is useful for specialists, clinicians who need to access it for action. There’s also a need for a unified report for other purposes, interdepartmental quality control, billing, et cetera. And we need versions that help the people who need to access that data, and often that’s in an EHR, which is difficult because we have to go from a pathology report to an EHR, depending on the system we have. Can you comment on that paradox?
Dr. Sinard (Yale): With respect to the discrete data, no matter what capabilities one builds into the anatomic pathology information system, one has to keep in mind the ability of the EHR to consume that information in one shape or another. Most anatomic pathology information systems have synoptic reporting with discrete data elements, yet most EHRs are not able to accept that data as discrete data, so there’s a significant loss along the way. Until that becomes an important element on the receiving side, there’s only so much one can do.
In relation to the paradox, there is a need for both. There have been a number of proposals over decades to create customized reports, different versions of reports for different recipients—oncologists, surgeons, patients. There’s a lot of overhead associated with that. What makes one version right for a particular subgroup is a dynamic thing.
One also has to consider practical aspects of how to update a report when new, additional information becomes available. How do you perpetuate that through the versions of the report and make sure the right people are notified about the additional or changed information? It is a big problem and not one that will be solved purely within the anatomic pathology information system.
Dr. Rashidi (UPMC): Many systems have their pathology whole slide image data in one data storage setup and the EHR under a different one, say, enterprise analytics. For us to be in the multimodal transformation space, it’s important for pathology groups to work closely with EHR central enterprise people to make sure they consolidate things and so initial infrastructure crosstalk becomes available, especially as you build out.
Another issue, not just for pathology reports but anything that’s a note, is our unstructured data. We have the opportunity now to incorporate generative AI in terms of converting some unstructured data into structured frameworks that can be easier to search and more accurate to search through. That would be helpful for future model building, digital pathology reporting, integrating various things and services that we do within pathology within the system, because then it becomes a framework that can be embedded similar to the EHR framework. That’s difficult to tackle because your conversion rate of unstructured to structured data is not 100 percent.
The future roadmap for everyone to want to tackle would be to have a strategy plan that’s hybrid so it incorporates a partnership of your vendor solutions and your homebrew solutions. Find vendors that are technically capable in, say, Epic Beaker integration, and have on-prem, cloud-based resource options. The vendor’s and your partners’ strategy plan should align with your institution’s strategy plan.
Lastly, people need to be aware that setting up a relationship between institutions or institutions and vendors is time-consuming, so it’s important they are as stable as possible. Make sure you understand the companies you’re interacting with and their financial stability, responsiveness post-deployment, and so forth. That’s in our best interest because we’re putting something in place for the long term.