Pushing the Envelope of AI Innovation in the Lab

Pathologist leaders say the future is evolving forward already in the clinical laboratory, though with a whole lot of asterisks. 

As an article recently published by Mayo Clinic Laboratories notes, “Artificial intelligence (AI) is reshaping laboratory medicine, not as a distant promise, but as a practical tool improving diagnostics, workflows, and quality.” And the article notes that “One of the most immediate impacts of AI in lab medicine is in workflow transformation. Tasks that once required hours of manual review, such as flow cytometry analysis, can now be completed in minutes with greater consistency and fewer errors.” What’s more, “Modern laboratory medicine is increasingly data-rich, with results spanning molecular diagnostics, histology, microbiology, and more.” 

A team of researchers at UPMC in Pittsburgh last year published a series of articles on what this will mean for pathologists and laboratory professionals. Matthew G. Hanna, MD, Liron Pantanowitz, MD, and their colleagues authored the series, whose first article was entitled “Future of Artificial Intelligence—Machine Learning Trends in Pathology and Medicine.” 

In that article, the authors write that machine-learning (ML) “platforms typically integrate a suite of tools and services designed to streamline the entire ML life cycle, encompassing data preparation, model training, validation, deployment, integration, monitoring, and feedback. [A]rtificial intelligence (AI) platforms… are revolutionizing medical imaging analysis and interpretation. The AI-ML platforms enable deployment of ML applications such as (semi)automated analysis of medical images (e.g., whole-slide images [WSI], dermoscopy, ophthalmology, X-rays, computed tomography scans, magnetic resonance imaging scans, etc.) to assist with detecting abnormalities, diagnosing diseases, and also predicting normal or benign conditions.” 

On the leading edge here is this: “[C]ombining histopathologic images with genomic data (i.e., DNA or RNA) can provide further insights into the underlying disease progression by incorporating such molecular mechanisms that can ultimately lead to a more precise diagnosis and prognostic measure.” 

Dr. Pantanowitz, who is chair and professor of Pathology at the University of Pittsburgh and a practicing pathologist at UPMC, notes that, “What has been done to date has been a series of very narrow tasks: taking pathology data in the form of image, text, or tabular form, and then making a diagnosis or prediction. That involves a narrow, simple path; but that’s not how medicine is practiced in the real world. A lot of data pops up in patient care that has to be managed. A real-world example would be a tumor board,” he says “If someone has a complicated cancer case, a tumor board would be held, and multiple specialists, including a radiologist and an oncologist, would be involved. The key is that, with the data, the best decision can be made, but that’s complex. So we’re trying to take all the important pieces of data in the clinical context, and combine them to make decisions. The molecular result, the clinical context, how have they responded to certain therapies, diagnostic images? Not just pathology images. So now pretend you’re at a tumor board, how would you, the tool, best manage the multimodal data? That’s the challenge—and the opportunity.” 

Per that, Dr. Hanna, who is associate professor and vice chair of Pathology Informatics at UPMC, emphasizes that “Everything in medicine is meant to be built on everything else. The premise is that, with multimodal platforms, the digital slide data of that patient’s specimen as well as the lab instrument values coming off the analyzers, and the genomic data coming off the PCR [polymerase chain reaction] instruments can be layered together. The whole premise is that it’s additive information that will help better diagnose and will predict something about that patient’s prognosis. In other words, the whole is greater than the sum of the parts.” 

Learnings early on in a long journey ahead 

What’s been learned so far in this journey?  

“A lot of things,” Dr. Pantanowitz emphasizes. At the beginning, he explains, pathologists were using AI as a second read. “But they realized they were missing out on all the efficiency gains, so they changed their minds. Now, when they get to work in the morning, their cases have already been screened by AI; so we’ve learned that. We’ve also learned that some of the algorithms couldn’t easily be generalized to all populations; we were biased based on training models only using our own specific population here. And lastly, we initially weren’t allowing our residents and fellows to use the AI; but we’ve changed that, and it’s been very beneficial.” 

Dr. Pantanowitz adds, “It turns out that currently each AI solution is very individual; there’s no standard way to apply AI.” And then there is the reimbursement model: do you buy it, or pay for clicks? “Unfortunately,” he says, “there’s no standard way of getting this reimbursed.”  

Both Dr. Pantanowitz and Dr. Hanna see a gradual forward evolution taking place; no one is going to wave a magic wand and have everyone in clinical laboratories making consistent use of abundant AI-based tools anytime soon, they emphasize.  

“A few things need to happen first,” Dr. Hanna says. There are three main categories of AI:  Image-based, text-based (OpenAI, ChatGPT), and tabular data, which includes the chemistry tests, blood counts, differentials. “AI can cover all three of those, and a lot of work in pathology has now also been done on the anatomic pathology side, because we generate digital slides. For the imaging aspect, digital pathology is an enabling technology for AI. And in order for that to happen, we need to move forward on the digital pathology curve. As for image-based workflows, not just for anatomic pathology, but also for microbiology, SPEPs (serum protein electrophoresis), and cellular hematology we can anticipate incorporating AI applications. For those tools to come to fruition, we have to move forward on digitization, which is lagging behind because of lack of reimbursement and other barriers.” 

Meanwhile, what should pathologists and laboratory professionals be thinking about right now?  

“I don’t think physicians should be afraid of AI; they need to spend time with it to learn how to trust it,” Dr. Pantanowitz says. “Looking ahead, AI is unlikely to replace doctors, but doctors who use AI will be better equipped than those who do not. At the same time, we must think carefully about using AI safely and sensibly. The paradigm is changing. Physicians are committed to doing no harm, but now patient care involves not only the doctor, but also the laboratory, the hospital, and even the AI developer. That shared responsibility means we need to apply more than just common sense, put safeguards in place, and continuously monitor how AI is used as this new paradigm evolves.”