AI roundtable: hopes, hurdles, hype vs. reality

It’s a tool, Jason, you’re right. In my days with BioImagene, roughly 12 years ago—this was before Roche acquired BioImagene—we had used the phrase, “The algorithm is the kit.” Much the way Sun Microsystems said in the 1980s “the network is the computer.” The algorithm is the kit, and the more we give it the semblance of a kit, the more we make it feel like a kit, the more it will be used as a tool and accepted as one.

Dr. Fuchs, you’re right in the middle of this as a founder and chief scientific officer. What is your reaction to what you’re hearing around these notions of explicable AI and the effect on the profession of pathology?

Dr. Fuchs (MSK): Jason mentioned something that’s very important: These tools and assays we are building are enabling for pathologists. It’s not only for efficiency, and it’s not only for the pathologist, but it’s also for all of the oncologists. There’s a lot of work in predicting all kinds of information outcome, response, et cetera from H&E images, and that puts the pathologist much more in the center of the patient treatment process, as the main diagnostician. In my experience at least, the vast majority of pathologists see that future and embrace it. It’s not age dependent. Young and experienced pathologists are very much looking forward to the future where they can reason about more interesting things than counting Ki-67 and so forth. So this mortal threat is not there, especially because it is the extension of previous work in the field.

These models are not black boxes. They’re completely transparent. You can investigate every weight, every connection, every design choice. They are complicated boxes, but they are not black boxes. In contrast to a human brain, you can investigate biases, you can correct for biases, you can do all kinds of tests and validation of these algorithms. You cannot do that with human experts.

Regarding explainability, we have been through that many times in machine learning. Just looking back at random forecasts, when they came out nobody wanted to use them because they’re so complicated compared with logistic regression. And why would anybody do something non-linear if you can do a logistic regression? And there were many groups that worked on visualization, explainability, and today it’s an off-the-shelf tool. The regulatory bodies pause not even for a second if you come up with that.

We’re seeing the same thing in deep learning. There are many groups, specifically those in pathology, that focus on explainability, and on the commercial side there’s not a single company that would advertise the replacement of pathologists. So these are tools, and if the tools are used the right way, they’re enormously empowering for pathologists, especially for pathologists in community practices, in large labs, and in the rest of the world. If you look beyond the ivory towers we have the luxury to live in, these tools will help patients in countries with health care disparities, in areas where the resources are not as plentiful as at Memorial or at UPMC.

Dr. Hipp, companies like Genentech and AstraZeneca are still reporting that there’s resistance to understanding this new role of the biomarker as driving therapy and of the pathologist as being the ultimate diagnostician. We sometimes call this the “town versus gown disparity.” Can you comment on the disparity and attitudes among pathologists, particularly in the community, as being perhaps not quite so eager to fulfill the new role that Dr. Fuchs so properly outlined, which is the new role for pathologists in our world?

Dr. Hipp (AstraZeneca): I don’t see it from the AstraZeneca side as much as being in the field of digital pathology for the past nine years, going to residents’ forums, conferences, and other events.

How we communicate these messages is what’s important. You have to speak multiple languages when you talk to this audience. You have to speak to a pathologist in the context of being a pathologist and understanding what it’s like and what you go through and the pressures you’re under. But you also have to use these quantitative terminologies—for example, explaining complex engineering where few pathologists have an engineering background.

C. P. Snow talked about the differences between humanities and science and speaking to the different worlds and languages. It comes down to communication and how I’ve been viewing and explaining things. Pathologists would always come up to me and say, “Oh, you’re trying to put me out of a job.” And I would explain, “No. This is going to make us better.” Then I would feel like I got a lot of buy-in and support and understanding.

That’s how we can communicate these messages and make more pathologists advocates for this technology—speaking to them as pathologists but yet understanding the science. More pathology informatics fellowships are crucial for the field as we begin to address this and to serve as the bridge between two different disciplines that never talked to each other before. And I’d love to hear the group’s thoughts on that.

Let’s start with Dr. Becich.

Dr. Becich (UPMC): This is a complex set of issues, but in pathology it should start with our member organizations, and, unfortunately, pathology’s member organizations are about as fragmented as they’ve ever been. The largest with a lobbying arm and the most financially successful is the CAP, and the CAP has begun to give a voice to what we need to do to engage community pathologists.

How we train our pathologists is where we’re not doing our job. To me, those are two significant choke points for infusing AI into pathology practice. The third choke point is with our LIS vendors, largely asleep at the wheel today in informing how we, in real time, start to infuse these technologies into pathology practice. Until those three choke points are alleviated, we’ll be a brute-force academic and industry innovation partnership to change it by disrupting this discipline.

Because of genomics, whole slide imaging, and the rise of microbiomics, particularly in the era now with infectious disease affecting us all, we can no longer stand behind the stovepipe industry of partners we have. And we have to transform our leadership in pathology practice from understanding how the community is informed by what’s happening at academic centers like Memorial and Pittsburgh and Michigan and Ohio State and other places that are now trying hard to push this gigantic ball up the hill to transform the culture of pathology practice.

Lisa-Jean, what’s your reaction?

Lisa-Jean Clifford (Gestalt): I agree with the need to be able to speak the correct language. I hear a similar story from many of the pathologists I work with in terms of embracing technology now and into the future. And age doesn’t matter—we are seeing pathologists of all ages and specialties becoming engaged with, and embracing, digital pathology.

What I am hearing are concerns over the use of AI. Some of the concern, and I see this more from pathologists at regional laboratories, is that AI will replace them. But what I’m hearing more often relates to the black box comment that was made. I agree that those of us who understand the technology know it is not a black box and that algorithms are fully transparent and well documented. The comment that resonated with me was how we have to speak the correct language to pathologists so that they understand and are comfortable with the fact that algorithms are locked once they are validated. I have had more than one pathologist state that they thought algorithms would continue to grow, evolve, and learn on their own indefinitely. The concern was that the algorithm used this minute isn’t going to be the same as the one used in 10 minutes, for example. It all goes back to speaking that correct language to the pathologist.

Our knowledge as the developers of the technology and of how it is used in practice needs to be communicated to the pathologists and the adopting laboratories in a language they understand and are comfortable with.

Esther, what’s your reaction to what you’ve heard? And the challenges of acceptance and alleviation of untoward anxiety?

Esther Abels (PathAI): It started when I was with Phillips when we got the imaging device cleared by FDA. We thought true clinical adoption was going to happen as soon as the first device, which was already being used in the clinical setting in Europe, would get to the market in the U.S. Then we got the clearance from the FDA, but still, the true clinical adoption at scale was slower than we all anticipated. Now looking at that, and when talking with users, pathologists, and vendors, for me what sticks out as a challenge comes back to interoperability.

In Europe, adoption is a little ahead of the curve compared with what’s happening in the U.S. We can learn from them. We can also learn from aerospace and automotive because there has been a lot of adoption and artificial intelligence usage already. Europe is ahead because its interoperability is better than in the U.S. I agree the LIS vendors could step up, but I also think the government can step up and make interoperability a regulation of sorts. I’m not in favor of adding regulations to health care, but I am in favor of guiding it or regulating it so that it becomes easier. We have to overcome all these hurdles to move this technology forward.

Dr. Singh, one of the concepts you talked about in your 2018 presentation at the Executive War College was the importance for innovation of what you call “cognitive diversity” or “transdisciplinarity.” Do you believe that pathology as you know it or experienced it has adequate transdisciplinarity?

Dr. Singh (Artiman): We can certainly learn from nature. If we appreciate the diversity in the Galapagos Islands and reason for a bit where it comes from, it’s obvious the importance of biodiversity there and how it leads to nature’s platform of innovation, which is life. Our human-made innovation is no different. It also requires diversity, among other parameters, and in this case it’s cognitive diversity. Jason made the point of how few full-fledged engineers have taken on pathology as a profession. That kind of answers the question.

There is a hierarchy of transdisciplinarity and I want to call out a couple such levels in this hierarchy. One is the transdisciplinarity across fields from engineering, AI, pathology per se, cell biology. There’s a transdisciplinarity at that level. Then there’s a transdisciplinarity at the integrative level because at the end we should not conceive the role as a pathologist only. Rather, we should conceive the role of a diagnostician, and as a diagnostician there has to be integration from multiple modalities of diagnosis, diagnostic tools, if you will. That’s the second or higher tier of transdisciplinarity that we need to incorporate. At Stanford we have a program on integrated diagnostics, and this is center stage in that subject.

On a scale of zero through 10, if I could summarize my comment, we are probably at a two or three in terms of transdisciplinarity.

Dr. Becich, where would you put us on a scale of one to 10 on the richness of our transdisciplinarity?

Dr. Becich (UPMC): Let me start with pathology in general. If 10 is perfect, pathologists are about a 0.9 in terms of their transdisciplinarity. That’s historical because pathologists generally get disrespected in medicine, often aren’t at the board tables when financial decisions are made, and aren’t necessarily the world’s best communicators. From the standpoint of academic pathology, most communications that go outbound to medicine do come from academic centers.

The largely discohesive nature of member organizations, as well as our member-facing organizations in pathology, leave us too fragmented to have a major impact and voice. Look at what’s happening with COVID right now. Pathology should be in the front and center in terms of understanding the importance of getting testing out there efficiently.

In academic health centers, diversity comes from who we put in our pipeline. We have a robust STEM pipeline to bring diverse and female voices to an issue that’s largely male dominated in technology and STEM. We need to retrain people in pathology. We need to embrace our communities, and if folks want to learn more about our approaches at Pittsburgh, I’m happy to talk about them.

Dr. Hipp, a final comment from you?

Dr. Hipp (AstraZeneca): I’d like to come back to this transdisciplinary concept in pathology. I’ve found it easier in industry to do multidisciplinary research, especially as a pathologist because I’m not on service. I don’t have to worry about grants. Often what I noticed is that in our health system we had the pathologists in the basement with their microscopes and the Center for

Engineering in a different part of the graduate school. So there’s that proximity divide. Whether I was in Silicon Valley or in the pharmaceutical industry, you sit in proximity. I sit next to the oncologists, the engineers, the bioinformaticians, and you need to have that kind of community. That idea where you don’t have to send an e-mail and wait for it to come back. You have that real-time discussion, that brainstorming.

Dr. Fuchs (MSK): I agree that interdisciplinary understanding and language and the willingness to get involved is more crucial than ever. But that’s true for all other disciplines these days so pathology is not unique in that sense. And in my experience, the vast majority of pathologists are eager to be part of the future and go forward boldly. There are always detractors and the detractors sometimes are loud. But we should not fall into false equivalency where we think that the fear of replacing pathologists is as big as the enthusiasm of the vast majority of pathologists who move forward to a better future and make the discipline much better.

Someone touched on education and that’s key. In Europe as well as in the U.S., pathologists must make pathology cool again. It’s important to present it as a modern and central discipline, which it is. I’m positive about the future.

Dr. Singh, how would you like to sum up your experience on the panel, your final thoughts, and if you might, what advice would you give for CAP TODAY readers in light of our roundtable?

Dr. Singh (Artiman): As always, anything that is philosophical, strategic, in discussion must end with something practical. So, at the end, what do we do? What’s the practical implementation? There are three pieces of general guidance.

Number one, digitize, digitize, digitize. Even if we have no idea how we’re going to use it, just digitize. Collect whatever data we can collect in our profession. If the incremental cost or effort of collecting one extra field of data is minimal, collect it, even if we have no idea today how it’s going to be used four, five, 10 years from now.

Number two, index, index, index. Whatever we can do in terms of creating some sort of a meta structure on top of data, do it. Even if you don’t have a program today in AI in a certain institution, just do whatever indexing is possible.

Third, annotate, annotate, annotate. Metadata is critical, and that means physical annotations by experts today, namely pathologists, is critical.

Even if we have no program in place, even if we have no idea how we’re going to use it, just making the data available in a digital form for your use and potentially others’ use in a syndicated, federated manner is going to push the field much further along than we can imagine.