By - September 05, 2023
Artificial Intelligence (AI), in which computer systems use machine learning and deep learning algorithms to mimic human-like tasks, is already changing the way pathologists and laboratory professionals work. While it is in its early adoption phases, at the rate at which the technology is advancing it will undoubtedly make huge leaps in coming years, radically transforming pathology practices. Pathologists who incorporate AI sooner than later are likely to find their practices becoming more efficient and effective, experts suggest.
“The scope of how AI can help pathologists is so broad,” says Dr. Carolyn Glass, MD, PhD, Co-Director of Duke University’s Division of Artificial Intelligence and Computational Pathology, an associate professor in the Department of Pathology and Division Chief of the Cardiovascular Pathology Division.
Dr. Glass explains that AI in pathology has already been used for years to assist with tasks like flow cytometry, such as peripheral blood smears, and CellaVision, assisting pathologists and laboratory professionals in counting cells. This will be the least of what AI can do. She breaks down the key areas where AI is, or soon will be, supporting pathologists:
Even the most expert pathologists disagree about difficult diagnoses, Dr. Glass says. “So we have AI algorithms for those use cases where it’s just purely diagnostic with high incidence of variability that we just can’t do based on the human eye.”
Indeed, according to the National Academy of Medicine,1 “AI approaches, specifically machine learning (ML), are especially well suited to the problems of clinical diagnosis, shortening the time for disease detection, diagnostic accuracy, and reducing medical errors.”
Triaging negative cases
Another area where AI could be extremely helpful is in high volume surgical pathology cases for such things as prostate biopsies or cytology specimens where, the majority of the time, they are likely to be negative. “An algorithm for a high negative predictive value is easily done. So that’s a big use case for saving time and money and resources, and something pathologists wouldn’t mind giving up,” Dr. Glass explains.
When it comes to treating patients with cancer, for example, precision medicine is evolving to be able to target the molecular differences in people’s cancers and create medication regimes specific to their cancer, Dr. Glass says. AI algorithms will be created to predict patient responses to drugs. “That’s going to have a huge patient impact not just for pathologists, but medicine in general. A lot of funding is going there, and it should.”
AI algorithms may also be able to bypass complicated, time-consuming molecular assays for complicated kinds of cancers, such as those with rare genetic translocations, Dr. Glass explains. “Machine learning algorithms are being developed that attempt to predict, just based on tissue and looking at the morphological features, if the tumor has that specific molecular alteration.”
In her own work, Dr. Glass has been part of an AI project at Duke that came up with the first algorithm to diagnose heart rejection using machine learning.
AI also has the ability to make pathologists’ jobs more accurate and efficient, according to Asa Rubin, MD, a pathologist and Medical Director of Pramana, an AI-enabled health tech company based in Boston that is modernizing the pathology sector. Dr. Rubin says that AI in pathology can be most effective in two ways: AI that helps humans do their work faster, more reliably, accurately and efficiently, and AI that does things humans can’t do.
He believes that AI’s ability to improve speed and accuracy will push more pathology practices toward digital pathology. “When examining tumors, pathologists look for specific characteristics that could mean more aggressive behavior. For example, lymphovascular invasion, which is where the tumor invades into blood vessels, can lead to metastases and is very important to document. It can also be very difficult to see,” said Dr. Rubin “People are going to start seeing the true benefits of AI when they discover its ability to catch those really small details that you don’t want to miss.”
Another of the ways AI can support pathologists is in a “retrospective” approach, where the AI flags any case where there is a disagreement between what the pathologist saw and what the AI saw, which, “Is really helpful for patient safety,” according to Dan Lambert, CEO of PathologyWatch. “If you run retrospective for long enough, then it helps the pathologist by surfacing what the AI found that's different,” Mr. Lambert says. "This also helps in the research context of finding where there are disagreements between humans and the AI." The pathologist can then have the AI’s data at their fingertips when they need it most.
Additionally, “Where AI really shines is when you’re looking for tiny bits of tumors on slides, for example,” Mr. Lambert says. “AI is also amazing at parsing large amounts of images that humans could miss if they’re going so fast. AI misses different things than humans do, so together, they’re pretty much perfect in combination.”
AI is already largely at play in the realm of digital pathology, says Massachusetts-based pathologist, Andy Beck, MD PhD, CEO of PathAI, an AI-powered pathology technology company. “Pathology is an area of medicine where the accurate and reproducible interpretation of images is really critical. There’s tremendous potential to use this technology to augment pathologists and make them more accurate and have even more information at their disposal in making diagnoses and guiding clinical care,” he says.
One example of how AI is already changing pathology is in the interpretation of liver biopsies. He explains that 25% of the U.S. population is considered at risk for non-alcoholic fatty liver disease, and thus biopsies are commonly used (particularly in clinical trials) to evaluate the severity of the disease. Pathologists must analyze key components including amount of inflammation in the liver, hepatocytes, and degree of fibrosis, to create a severity score, a difficult thing to do without assistance.
“Instead of a pathologist acting alone, the next step after staining is to put these slides into a whole slide imager to create large image files, to process those slides with an AI system whose job it is to do very similar what you might ask a resident or trainee to do.” The end result is that rather than having the pathologist count cells themselves, the AI would aggregate that information for them, allowing them to look at it, synthesize the data, and either agree with the AI or change it if they disagree.
Another way that AI will help is in evaluating tumor biomarkers, says Dr. Rubin. AI has already been shown to be capable of predicting the expression of a specific cancer biomarker known as PD-L1, a common tumor biomarker for which there is variability in how it’s scored and interpreted by pathologists.
One AI algorithm was able to accurately predict PD-L1 expression in 70% of tested cases just by looking at the tumor morphologically. This proves that there is information contained simply within the histology of a tumor that humans cannot see but computers can detect and translate into meaningful outcomes. “So, there’s all kinds of things that open the door to us understanding or treating tumors better via AI algorithms,” Dr. Rubin explains.
While every laboratory and pathology practice can begin to implement AI tools, AI may be more useful for laboratories that have a specialty focus, Mr. Lambert says. “If you’re a focused laboratory, say all you do is dermpath, or just prostate [biopsies], the value proposition and gain from using AI is really clear. But if you’re a lab that does 5,000 things, it’s not going to be as effective—you’re going to have to wait until some of those things catch up.”
Additionally, having older or outdated laboratory information systems (LIS) may make it harder to interface with newer AI technology, he says. So, laboratories may be looking at a lot of financial output up front for long term gain. However, LIS are not cheap, and Dr. Glass explains that digital scanners can run upwards of $300,000. Changing technology means looking closely at a laboratory’s budget.
“No matter how you splice it, you’re going to get so much efficiency and safety out of [upgrading technology] that it’s worth making the investment,” Mr. Lambert says.
While AI is unlikely to replace pathologists, Mr. Lambert does feel that pathologists who incorporate AI will outlast those who don’t.
Adler-Milstein, Julia et. al. “Meeting the Moment: Addressing Barriers and Facilitating Clinical Adoption of Artificial Intelligence in Medical Diagnosis.” National Academy of Medicine. https://nam.edu/meeting-the-moment-addressing-barriers-and-facilitating-clinical-adoption-of-artificial-intelligence-in-medical-diagnosis/