Is the infrastructure surrounding AI-augmented digital pathology evolving quickly enough to support AI safely and sustainably at scale?

In this roundtable, four experts spanning clinical pathology, digital pathology, and AI development examine where the real barriers lie and how they impact laboratories. 

Their discussion reveals a field at an inflection point: AI enhancements are driving digital pathology adoption, yet successful deployment depends as much on workflows, partnerships, and economics as it does on algorithm performance. And with AI-only biomarkers potentially turning digitization from an operational choice into a clinical necessity, the pressure to close those gaps may soon intensify.

In your experience, is current laboratory infrastructure prepared for widespread AI integration?

Dylan Miller: My experience with AI is primarily with machine learning and deep learning algorithms applied to scanned pathology slides. In that context, I would say we're not yet ready for widespread AI integration, although we're getting closer every day.

One of the biggest infrastructure obstacles is that AI requires a digital version of the slide. In the United States, most laboratories are not yet scanning slides at the scale needed for widespread adoption of digital pathology, let alone AI. A major factor is the sheer volume of data these images generate and the associated storage requirements. We simply haven't progressed far enough along that curve for large-scale digital pathology to be a practical reality for most hospital systems and independent laboratories.

Beyond the technical infrastructure, there are also significant regulatory and economic considerations that underpin AI adoption. 

Lisa-Jean Clifford: I agree with Dr. Miller. Overall adoption of digital pathology remains relatively low, and AI adoption represents an even smaller subset. We’re also hearing concerns around governance and cost that can present additional barriers.

That said, among our customers who have adopted digital pathology, a significant percentage are deploying AI within those environments – for research, education, and primary diagnosis. What’s particularly interesting is that once customers adopt their first algorithm, they often add others very quickly because they see the value these tools can bring to clinical practice.

Marco Comianos: While much of the pathology market is not yet digital, many academic medical centers have already evaluated digital solutions and AI – and, in some cases, developed their own AI tools. They have taken many of the foundational steps needed for adoption.

There are also relatively low-risk applications of AI that can provide good starting points for hospitals. The important thing is to begin gaining experience with the technology. I compare it to the autonomous driving features in cars: when you first use them, you’re not necessarily comfortable trusting them. You need to see how they perform against your own driving skills and understand their limitations. Over time, you become more comfortable with what they can and cannot do.

Healthcare organizations need to go through a similar learning curve with AI – using and validating these tools, understanding where they add value, and identifying the risks that need to be mitigated.

How urgently do we need to address the infrastructure gaps for AI adoption?

Felix Faber: There may be forces in the market that require us to move much faster – AI-only biomarkers are one example. If an AI-based biomarker becomes necessary to determine eligibility for a particular therapy, the current rate of digitization simply may not be fast enough to meet that need.

I don’t think the existing infrastructure can support that kind of rapid adoption as it stands today. We may need different deployment models – for example, smaller scanners with AI capabilities built in, or direct connections between existing scanners and AI platforms. In that sense, the emergence of AI-only biomarkers could itself become a driver for accelerating digital pathology infrastructure.

Dylan Miller: I’ve been following digital pathology for a long time, and I’ve seen a significant shift in what’s driving adoption. Initially, the case for going digital centered largely on remote work – being able to sign out cases from a distance, access subspecialty expertise in another location, and improve workload balancing and efficiency.

Increasingly, though, that conversation has shifted toward AI. In my view, AI has become the number-one driver for adopting digital pathology. Those other benefits remain important, but more and more, I’m hearing institutions say they need to go digital specifically so they can take advantage of AI tools.

Lisa-Jean Clifford: I agree. And while overall adoption remains relatively low, it is accelerating significantly. Adoption rates have roughly doubled over the past 12 to 18 months, and almost every large and midsize organization – whether an integrated health network, hospital, or independent laboratory – now has some form of digital pathology evaluation initiative underway.

So although the starting point is relatively low, the trajectory is clear: digital pathology adoption is moving sharply upward.

Are fragmented digital ecosystems becoming the biggest obstacle to AI at scale?

Lisa-Jean Clifford: From the image management system and interoperability perspective, I don’t see fragmentation as an insurmountable barrier. Several IMS vendors can integrate AI directly into their platforms, bringing it into the pathologist’s existing workflow. That is an important driver of adoption and can also help address governance and safety concerns. Rather than moving between multiple applications, the pathologist can be confident that the AI overlay corresponds to the same patient and image they are reviewing – all within a single environment.

There are certainly systems with less open architectures, and those vendors will either need to re-architect their platforms or risk being left behind. As Dr. Miller noted, AI is increasingly the impetus for going digital.

And this extends well beyond diagnostic AI. We’re seeing emerging applications in quality control, reporting, education, clinical trial matching, and many other areas. The breadth of those applications makes interoperability increasingly important as AI becomes more embedded in pathology and healthcare.

Marco Comianos: Fragmentation has been a problem, and to some extent it still is, but I think the market will naturally correct it. Customers want the freedom to use whichever applications best meet their needs, so it’s in the interests of IMS vendors to make their platforms interoperable with different solutions.

The technology is moving so quickly that the best solution today may not be the best solution tomorrow. Laboratories need the flexibility to interchange products and adopt best-of-breed technologies as they emerge. I think the market is already moving in that direction, with interoperability increasingly becoming an expectation rather than an option.

Felix Faber: There are certainly examples where interoperability works well, but there are also environments where fragmented IT infrastructure remains a significant challenge. When you’re deploying AI globally and at scale, you inevitably encounter a wide range of IT issues.

That said, I wouldn’t consider fragmentation the biggest obstacle to AI adoption. In my view, the lack of reimbursement is the more significant barrier. Interoperability and IT infrastructure are important, but I would probably rank them second or third among the challenges we need to solve.

Dylan Miller: I think about this from the perspective of the traditional glass-slide workflow. Pathologists have spent decades optimizing that process. When we sit down to sign out cases, everything is designed for efficiency: we scan the slide barcode to bring up the case in the LIS, our dictation tools are ready, and voice recognition is integrated into the workflow. That level of efficiency sets the benchmark that digital pathology has to meet.

An AI-enabled digital workflow needs to be just as seamless. Ideally, the pathologist should be working within a single screen or interface rather than moving back and forth between multiple applications. Every additional window or system creates an interruption and, potentially, an opportunity for error or a case mix-up.

It’s encouraging to see vendors recognizing that and working toward interoperability. For digital pathology and AI to succeed in routine practice, they have to deliver that same seamless, integrated sign-out experience. If the digital workflow isn’t at least as efficient as glass, widespread adoption is going to be very difficult.

What constitutes evidence that an AI tool is safe and effective?

Felix Faber: Safety and effectiveness start with a strong body of published evidence and robust validation data. From a regulatory perspective, the product should either have the appropriate FDA clearance or be implemented through the applicable laboratory-developed test pathway.

But even strong external validation isn’t enough. Laboratories should perform their own validation using their own data before adopting an AI tool. You need to understand how that product performs in your specific setting rather than relying solely on results generated elsewhere.

Marco Comianos: We often talk about algorithm drift, but it’s not necessarily the algorithm itself that changes – it’s the environment around it. New scanners are introduced, software is updated, image formats change, and staining can vary. An AI solution trained on a large and diverse dataset may generalize well across many of those variables, but you can’t anticipate everything.

That’s why validation shouldn’t be viewed as a one-time exercise. Having appropriate validation datasets and routinely evaluating performance as the operating environment changes are key to ensuring an AI tool remains safe and effective.

Lisa-Jean Clifford: I would also emphasize the importance of understanding the data used to train and validate the algorithm. Vendors should be transparent about the breadth and volume of that data: how many scanners and scanner models were represented, which tissue types were included, how many images were evaluated, and, where relevant, the number of positive and negative cases.

That information does more than provide confidence in the algorithm’s performance. It helps laboratories understand the boundaries of its intended use – where the tool has been adequately validated and, just as importantly, where it should not be applied.

Marco Comianos: Another important element is maintaining a close partnership between the laboratory and the AI vendor. There is considerable nuance in how pathologists interpret cases, including interobserver and intraobserver variability, and AI cannot account for every variable on its own.

That relationship needs to continue after implementation. If something in the environment changes – such as a new scanner or a software update – the laboratory and vendor should work together to assess the impact and test the AI appropriately before those changes affect patient care.

Dylan Miller: “Safe and effective” is an interesting standard in this context because it comes from the clinical trial and FDA paradigm. Clinical trials may be designed around non-inferiority rather than superiority – the goal is often to demonstrate that a new approach is no worse than the established one.

But applying that framework to digital pathology and AI raises an interesting question: Where is the evidence that the way a pathologist reads a glass slide is itself “safe and effective”? We don’t have clinical trials validating that practice; it is grounded in decades of experience. Yet human interpretation of glass or digital slides remains the comparator against which AI is evaluated. In that sense, we’re working with an imperfect gold standard, particularly given the interobserver variability inherent in pathology.

I think the best approach is to surround AI with the same kind of quality framework we apply elsewhere in laboratory medicine, considering preanalytic, analytic, and postanalytic factors.

It’s also important to recognize that “AI” in digital pathology encompasses very different types of tools. At one end are quantitative algorithms that count events or findings. Then there are pattern-detection applications, such as estimating tumor cellularity or identifying rare foci of cancer. Increasingly, we’re also seeing algorithms that use H&E images to predict molecular alterations, patient outcomes, or response to therapy.

The further we move along that spectrum, the more difficult it becomes to define the gold standard. A counting algorithm is relatively straightforward: I can count something, the computer can count it, and we can compare the results. Cancer detection introduces more interobserver variability, but consensus approaches can help address that. With outcome-prediction algorithms, however, establishing reproducibility – and demonstrating that the result is truly safe and effective – becomes much more challenging. I think that will be one of the major issues we have to confront.

As AI becomes embedded in workflows, who is accountable when algorithms influence clinical decisions?

Lisa-Jean Clifford: Ultimately, the pathologist is responsible for the patient’s final diagnosis. The pathologist can accept, reject, or modify the information provided by an AI algorithm as part of their interpretation.

But accountability for the technology and its integration is shared across the vendors involved. The relationship between the digital pathology or IMS vendor and the AI vendor has to be very closely aligned, with clearly defined responsibilities for how their applications interact.

For example, the IMS vendor has a responsibility to ensure that the results received from the AI algorithm are displayed accurately and without alteration. These responsibilities need to be clearly established contractually, but they also require a close operational partnership between the vendors.

Felix Faber: I agree with Lisa-Jean, and I would add that direct communication between the laboratory and the AI provider is extremely important, while of course keeping the IMS provider involved.

In our experience, when all communication has to pass through a single point of contact, the process slows down and important details can get lost in translation. The best outcomes come when the customer can communicate directly with the AI provider while keeping all parties informed. That creates a much more effective partnership when questions or issues arise.

Marco Comianos: I largely agree. When it comes to scaling AI, the foundation you put in place matters. Organizations need clear SOPs and governance around how these tools are going to be used before they begin deploying them more broadly.

If you have the right processes in place, they can scale along with the technology. If you don’t, the problems will scale as well. That means clearly defining and communicating how AI tools should be used within the organization, while working closely with vendors to ensure the technology is implemented correctly within the clinical workflow.

Dylan Miller: I agree with what’s been said. At least in the United States, clinical applications of AI currently operate with a “human in the loop.” Ultimately, a pathologist is overseeing the process and making the decision to finalize the report, so the responsibility – and ultimately the liability – rests with that pathologist.

If we reach a point where AI is deployed without a human in the loop in certain settings, this question becomes much more complicated. But for now, the pathologist remains accountable for the final clinical decision.

As one of those pathologists, I also believe that responsibility extends to understanding and validating the AI tools we use. When I deploy an AI algorithm, I want to be confident that the slides generated in our laboratory are comparable to the data on which that algorithm was trained. There are many preanalytic and technical variables that can affect performance, including scan quality, image format, and differences in how scanners reproduce color.

That’s why rigorous local validation is essential. We need to demonstrate that slides produced in our own laboratory perform as intended with the algorithm, rather than assuming that performance demonstrated elsewhere will automatically translate to our environment.

Look out for part two of the discussion to hear the key considerations for sustainable AI workflows.

Newsletters

Receive the latest pathologist news, personalities, education, and career development – weekly to your inbox.

I have read and understand the Privacy Notice*

Subscribe