Pathologist-computational scientist collaboration: The why, when, and how
September 2022—Do you want to play a role in solving a problem or improving a process in your laboratory via machine learning or artificial intelligence but don’t know where to begin? If so, take some time to learn, listen, share, and, perhaps, have a cup of coffee, says Scott Doyle, PhD, biomedical engineer at the State University of New York at Buffalo.
Dr. Doyle, who is also associate professor in the Department of Pathology and Anatomical Sciences at SUNY-Buffalo, is part of a growing trend of computational scientists and pathologists working in concert to apply computational methods to improving patient care. He is currently working with pathologists on an AI-based system that uses image analytics to predict oral cavity cancer recurrence risk.
Yet, while interest in machine learning, computer vision, and other AI tools continues to grow, pathologists, by and large, don’t know how to turn their ideas for creating and employing such tools into real-world projects that would benefit patient care. In the Cliffs Notes version of how to undertake a pathologist-computational scientist collaboration, Dr. Doyle says to familiarize yourself with the technology by listening to talks and reading textbooks and papers. “And if you think it’s a tool that you can put to use, find someone who does that type of research and take them out for coffee and ask them questions. Present your ideas. I think that’s probably the best way to get started.”
Michael Feldman, MD, PhD, professor and vice chair of the Department of Pathology and Laboratory Medicine at the University of Pennsylvania, agrees. Dr. Feldman, who has participated in such collaborations for more than 20 years, recommends that pathologists develop a basic understanding of AI, focusing on machine learning, before considering a potential project with a computational scientist. He advises enrolling in the annual Association for Pathology Informatics digital pathology and AI workshop (coming up in November). Then, if you think you have a worthwhile project, he says, “get out from behind your microscope” and find collaborators. “Find out where the mathematicians and the computational folks hang out. They’re on the university side, not at the hospital. Go to your engineers. Go to your mathematicians. Go to your computer science folks. Find somebody who’s hungry.”
Because Dr. Doyle welcomes such collaborations, he developed a checklist to determine the feasibility of working with pathologists on proposed projects. He uses that standard set of questions to guide initial project meetings with potential collaborators. He starts by asking pathologists to describe the clinical problem they’re attempting to solve and the available data. “How many patients? How many cases? I’m big on data being the primary stuff AI is built upon,” he says. “The algorithms can be used for many things, and what differentiates them is the data set.” Dr. Doyle also asks if data were collected from patients during treatment. “If you’re interested in an AI project but the data are expensive to obtain or not routinely collected from your patients, it means the cost of implementing such a system goes up,” he explains. But if the data are collected routinely, “it becomes easier to then collect more of that data,” and the AI project or application would likely also fit the clinical workflow.
During that initial meeting, pathologists should describe the disease’s current standard of care, he says, including the types of tests used during diagnosis and treatment and the tests’ percentage of predictive power, as well as how they hope to use machine learning and AI to improve the current process. “As an engineer, I have a limited understanding of how patients are managed and what the decision points are,” he says. “The earlier you can show me what it is you want the computer to do, the sooner I’ll have an idea of what kind of tools in my tool set can be applied to this problem.”
Dovetailing with this initial information-gathering step is defining the project’s clinical impact. “If your system is successful, all your hypotheses are supported, and all your data is significant,” says Dr. Doyle, “what’s going to happen to the patient who walks into your clinic tomorrow that isn’t going to happen if they walk in today? And how many people is that going to affect?
“AI can do whatever you ask it to,” he continues. “It can detect blood cells or measure the size of the tumor. But if the size of the tumor doesn’t change what happens to the patient, then it doesn’t have an impact on how the patient is managed.” Furthermore, problems that medical staff can solve quickly and easily should not be the focus of AI projects, says Dr. Doyle. “So we try to make sure we focus our energies on things where AI is going to improve the management and treatment of patients.”
Dr. Feldman takes this line of thought a step further, pointing out that pathologists “are sometimes quick to say, ‘Let’s just solve my problem,’ and it’s almost never just your problem.” You will get an answer to a problem with machine learning, he explains. However, machine-learning models should be robust and useable across health care systems. “What you want is an answer you can generalize. If it only works on my data set and it doesn’t work on your data set, what good is that for you?”
Another factor in determining whether to collaborate on an AI project is the ability of the machine-learning system to achieve the minimum level of acceptable performance accuracy, Dr. Doyle says. “What I ask is, ‘How accurate are you currently in analyzing or assessing these patients, and how accurate will the tool need to be in order to improve the way you treat such patients?’ For some tests, that number is quite low. Maybe only 60 percent of people will benefit from this, but that’s still valuable,” he says, if it improves the standard of care.
Dr. Doyle also cautions pathologists about the potential drawbacks of acquiring external data. He asks them, “Where is the data? And can I get it? And how hard is it to get?” because acquiring data from other institutions and transporting it between facilities can be a bureaucratic nightmare. He cites as an example the oral cavity cancer project in which he is participating. It took about a year to finalize the data-use agreements, he says. And each of the numerous sites that is contributing data required its own agreement. “Depending on the institution, it could take even longer [to get permission],” he notes. However, a more recent approach to training machine-learning models, known as federated learning, may ease some of the bureaucratic burden of transferring data, Dr. Doyle says.
Federated learning may also be the key to allowing pathologists from community hospitals and smaller institutions to participate in machine-learning projects, according to Drs. Doyle and Feldman. The method, pioneered by Google to improve text prediction in mobile keyboards, allows an institution to transfer its data model to another institution without moving the original institution’s primary data, thereby allowing the model to be trained on the second institution’s data. “The federation occurs as each individual model gets trained locally, and the model then aggregates back without moving any of the protected health information,” Dr. Feldman explains.
In this manner, federated learning has the potential to help “some incredibly smart folks practicing in the community and at smaller hospitals who have really good questions and really good ideas and just don’t know how to pull the right kind of team together to address them,” says Dr. Feldman. He emphasizes that good ideas can come from any location, and “there’s nothing special about academicians that gives them all the answers.”
Yet, project development and data networking aren’t the only considerations for those interested in working on a collaborative machine-learning project, according to Dr. Feldman. “It’s about building trust and understanding,” he says, noting that part of relationship building involves learning each other’s disciplines.
Dr. Feldman and biomedical engineer Anant Madabhushi, PhD, of Emory University, have long been collaborators and have spent years educating one another. “I know how to read the papers and understand the statistical analysis of the data,” says Dr. Feldman, “and I learned some of the language, mostly because he [Dr. Madabhushi] was patient and took the time to explain it to me in nonmathematical terms. And I taught him medicine. So I went halfway towards him and he went halfway towards me, and we start out every new project that way.”
“Don’t assume we know anything about cell biology or cancer even if we’re biomedical engineers,” Dr. Doyle adds, “because it could have been a long time since we learned about that. Being patient with our lack of understanding of the things you take for granted is important.”
Dr. Feldman’s collaborations have hinged on his ability to catalyze and connect practitioners of different disciplines, and he advises others to do the same. “Bring the cast of characters in medicine together through your collaborations and connections,” he says. Over the years, Dr. Feldman has introduced Dr. Madabhushi to several clinicians, widening the scope of projects they tackle. For example, he and Dr. Madabhushi are working on a project with cardiologists to predict cardiac transplant rejection and are also collaborating on a project with surgeons who specialize in head and neck cancer.
Tuck your egos away, Dr. Feldman advises, and bring every relevant specialist into the mix because in the end, “it’s just people solving problems. . . . We don’t always get it right,” he adds. “Then Dr. Madabhushi and I have a scotch and say, ‘That was a nice idea, but it didn’t quite pan out, so let’s try something else.’”
—Charna Albert
NovoPath and FrontRunnerHC forge reseller agreement
NovoPath has signed a reseller agreement with FrontRunnerHC that brings together NovoPath’s software-as-a-service–based laboratory information system, NovoPath 360, and FrontRunnerHC’s LabXchange and PatientRemedi solutions.
The FrontRunnerHC products, which seamlessly integrate with the NovoPath 360 platform, securely capture patient demographic, insurance, and financial information at test order entry and then cross-check that data and fix errors in real time. The systems then automatically transmit the data to the NovoPath 360 solution and trigger the proper workflow within the LIS. This allows physicians to quickly diagnose patients and communicate test results via email or SMS text, according to a press release from NovoPath.
NovoPath markets the NovoPath 360 platform to anatomic, clinical, and molecular pathology laboratories.
NovoPath, 732-329-3209
VistaPath and Gestalt offer combined AI solution
VistaPath has announced plans to combine its Sentinel automated tissue-grossing platform with Gestalt Diagnostics’ artificial intelligence requisition engine, or AIRE, an AI algorithm for accessioning.
Sentinel uses a video system designed to assess specimens and create a gross report faster and more accurately than manual processes. It continuously monitors the cassette, container, and tissue to reduce mislabeling and specimen mix-ups and retains original images for downstream review.
AIRE is an intuitive, scalable, continual-learning AI algorithm that detects requisitions and other types of forms. It interprets data and handwriting from scanned paper requisitions and sends electronic orders to the LIS.
Gestalt Diagnostics, 509-492-4912
Tribun Health updates telepathology system
Paris-based Tribun Health has redesigned its TeleSlide Patho 5 next-generation telepathology solution for remote case review and second opinions.
The revamped system sends SMS and email notifications to alert pathologists when they have been assigned a review request or when a response to a request has been received. Users can share documents via a drag-and-drop process or by clicking an “add document” button.
The home page of the updated platform features at-a-glance views of the user’s second opinion requests, second opinion assignments, and recently closed cases. Pathologists seeking a second opinion can use a drop-down list to select a reviewer based on the latter’s area of expertise. They can also use a drop-down menu to identify sites or organizations with which they would like to collaborate on case reviews.
Among other system advances is a simplified approach to adding and removing case reviewers and setting account passwords.
TeleSlide Patho 5 is available for research use only in the United States.
Tribun Health, +33-1-89-20-00-07
Paige and Sonora Quest enter digital pathology partnership
The computational pathology company Paige has announced that it will implement digital pathology workflows at Sonora Quest Laboratory sites across Arizona.
Paige will provide Sonora Quest with a custom installation of its suite of digital pathology software, including its FullFocus viewer and Paige Prostate Detect and Paige Breast systems.
Sonora Quest will share the resulting data with clinicians through an existing partnership with Pathology Specialists of Arizona, a Mesa-based physician practice.
Phoenix-based Sonora Quest is a joint venture between Banner Health and Quest Diagnostics.
Dr. Aller practices clinical informatics in Southern California. He can be reached at raller@usc.edu. Dennis Winsten is founder of Dennis Winsten & Associates, Healthcare Systems Consultants. He can be reached at dwinsten.az@gmail.com.