“AI is revolutionizing pathology,” says Bobbi Pritt, Chair of Clinical Microbiology at Mayo Clinic. We connected with Pritt after her presentation at the DP&AI: Europe congress to discuss AI in action in image-based clinical pathology – today and in the future.
From your perspective, what does AI currently do well in routine image-based pathology, and where is it already making genuine diagnostic impact?
AI performs especially well for repetitive, quantifiable tasks – the kinds of things people don’t necessarily want to do manually. Examples include counting mitotic figures or measuring a stroma-to-tumor ratio. In my area, it’s more focused on detecting microorganisms.
AI is also very useful for quality control and prescreening of stained or scanned slides – checking whether a slide is stained correctly, and flagging issues such as bubbles, folds, or artifacts. It would be helpful to have AI quickly scan slides before review.
Another strong use case is triage and workflow automation. AI can help prioritize slides by routing likely negative cases for rapid review and flagging abnormal cases for closer evaluation or referral to a specific pathologist.
In clinical microbiology, we mainly use AI to screen out negative slides, which represent about 90 to 95 percent of specimens. This has reduced review time from about five minutes per slide to roughly 30 seconds, creating a major efficiency gain and allowing us to focus more time on true positive cases.
AI may also contribute to malignancy detection, particularly as a second-read tool. That area looks promising and may expand further.
Many pathologists view AI as an exciting tool but aren’t sure how it fits into daily workflow. Can you describe practical, non-industry-driven use cases where AI has supported or enhanced diagnostic efficiency?
I don’t think we’ve reached “true” artificial intelligence yet. What we have today is closer to augmented intelligence – tools designed to support and extend human expertise rather than replace it.
This approach is especially important because we are facing a global shortage of trained technologists. Even when we are able to hire new staff, it can take six months or longer for them to become fully trained and confident. We want to use their time wisely and support retention – especially by reducing repetitive microscope work that can contribute to fatigue or ergonomic strain.
As a clinical pathologist overseeing the laboratory as medical director, this is directly relevant to pathology practice. AI is becoming a practical tool to help laboratories maintain quality and efficiency despite workforce constraints.
What criteria do you rely on when evaluating whether an AI tool is truly ready for clinical deployment versus still in the exploratory or research phase?
The first step is always to ask: what problem are we trying to solve? Sometimes a new technology looks impressive, and the instinct is to ask how we can use it. But it’s more effective to start with the clinical or operational need and then evaluate whether a tool actually addresses it.
Next, I look at validation and evidence. Has the tool been tested on independent, representative datasets – or only on data provided by the manufacturer? Is the evidence publicly available? Has it been peer reviewed and published, or is it limited to internal industry reporting?
I also consider robustness, bias, and safety. How does it perform across different real-world conditions – such as different scanners, monitors, and laboratory systems? Is it adaptable, or is it limited to one tightly controlled setup?
And finally, integration matters. Can the tool connect with the laboratory’s existing infrastructure, such as the laboratory information and image management systems, and can it fit into routine workflows? In practice, workflow integration is often the most important factor for clinical adoption.
If a tool does not meet these criteria, I view it as a research tool, not something ready for routine clinical use.
How can pathologists remain central to AI development and validation while ensuring that clinical needs – not technological trends – drive innovation?
There are several things we can do as pathologists. First, we need to own the problem definition. Our role is to clearly identify what problem we are trying to solve, because we understand where the real friction points are – whether it’s diagnostic ambiguity, complex grading systems, error-prone handoffs, or inefficiencies in workflow. We need to be involved early to ensure the technology is being developed for the right purpose.
As tools move into development and validation, pathologists should be at the table as leaders, ideally as principal investigators or co-investigators. That includes helping design the dataset and defining the “ground truth.” I say that with a caveat: ground truth is becoming harder to define, because some systems are now so strong that even traditional “gold standards” can be challenged. Still, it remains essential for pathologists to determine how ground truth will be established and how discrepancies will be adjudicated – for example, if the AI output differs from expert review.
We should also push for explainability when it matters. We don’t always need to understand every technical detail of how a system works, but we do need to ensure the outputs are interpretable and clinically meaningful. At minimum, we should understand how the model was trained and what data it was trained on.
Another key responsibility is training the next generation. AI needs to be integrated into pathology education so that trainees can learn how to use these tools appropriately and safely.
And crucially, pathologists must play a role in advocacy, locally, nationally, and globally. This includes discussions with payers and policymakers – especially in the US – because reimbursement will determine whether AI tools can be adopted sustainably. There is concern that increased efficiency could reduce payment per case, even as laboratories absorb the added cost of AI. Pathologists need to be part of those conversations to protect both quality and access.
As someone deeply involved in clinical microbiology, do you see particular areas within this discipline where AI could add significant value?
In microbiology, AI can be applied to a range of image-based tasks, such as detecting microorganisms, identifying rare events, and recognizing growth on culture plates.
But there are also many non–image-based applications that we’re only beginning to explore. One major area is interpreting complex datasets from metagenomics. The broader the testing approach, the more organisms we can potentially detect – fungi, bacteria, viruses, mycobacteria, parasites – but we also generate a large amount of additional information that can be difficult to interpret in context. AI may help us make sense of that signal, particularly when combined with host-response data, such as patterns suggesting a viral versus bacterial infection.
AI could also support risk prediction. For example, identifying patients with sepsis who are more likely to deteriorate or develop complications. Another emerging use case is linking phenotypic susceptibility patterns with genotypic resistance markers, helping to interpret antimicrobial resistance more efficiently.
AI can also help with operational optimization in the lab. This could include predicting culture positivity, improving workflows, and refining protocols – such as determining how long cultures truly need to be incubated based on historical outcome data.
Overall, AI in microbiology extends well beyond slide or plate imaging, and its role in data interpretation and workflow improvement is expanding.
How do you approach training residents and fellows so they understand both the capabilities and limitations of AI in digital pathology?
When we teach residents and fellows about AI, the first step is making sure they understand the foundational concepts. Most will not be writing code, but they should know the difference between a training set and a validation set, and why those datasets must be separate. They also need to understand how bias can be introduced and why that matters for clinical use.
Hands-on exposure is just as important. If our laboratory is using AI tools in routine work, trainees should be using them too, so they gain practical experience.
We also need to teach trainees how to read AI papers critically. The volume of AI-related publications has exploded, and not all algorithms are developed or validated to the same standard. Future laboratory leaders need to be able to evaluate the evidence and understand what is clinically meaningful.
We also reinforce the central role of the pathologist. AI is best viewed as augmented intelligence – a tool, like any other piece of laboratory technology. We need to know how to use it appropriately, but it does not replace clinical judgment.
One important risk is automation bias. Technologists or pathologists may be tempted to accept the AI output as the “truth,” rather than treating it as one input. We have to remind teams that they are trained professionals, and they should still evaluate cases independently. In practice, that means encouraging blinded review when appropriate and keeping a human in the loop, even when AI results appear confident.
Looking ahead, which trends in digital and computational pathology do you think will most influence diagnostic medicine in the coming decade?
There are several trends I’m excited about. One is end-to-end digital workflows, with more automation powered by highly accurate large language models. The goal is to pull together data from multiple sources so that when someone sits down to review or sign out a case, they already have everything they need in one place.
For example, even if a patient has 20 years of records, the relevant details could be distilled and presented upfront – such as immunocompromised status, a history of renal transplant, prior infections, the latest culture results, and recent imaging interpretations. Having that information available at the point of review could be extremely useful, as long as it is accurate and reliable.
I also think regulation, payer policies, and standards will become increasingly important. As more tools enter the market, government agencies, regulators, and payers will need to catch up to support safe and sustainable clinical adoption.
Another major area is multimodal models that combine images with clinical notes, lab values, genomics, and other text-based data. This ties back to the need for better integration of the medical record into diagnostic workflows.
Finally, I’m interested in augmented reporting. Pathology reports are typically written for clinicians, but it would be valuable to generate additional versions automatically – for example, written at a simpler reading level, translated into multiple languages, or tailored to different audiences. The technology to do this already exists, but it has not been widely implemented in routine practice.
Newsletters
Receive the latest pathologist news, personalities, education, and career development – weekly to your inbox.
