Lessons Learned, with Richard Levenson
More than just pigeons: Levenson has a vast and varied career in pathology, microscopy, and computational tools. Here, he shares his experience and his thoughts on the future
At a Glance
- Pathology is an excellent career choice for those who want to focus on research as well as clinical work
- Novel techniques that solve many of pathology’s logistical problems may be the way forward for clinical microscopy
- The field’s transition to digital is promising, but has yet to overcome significant challenges
- When digital pathology allows its practitioners to be free of the slide, rather than simply adding steps to its processing, its popularity may spread
Pathology has possibilities
I was doing reasonably well as a medical student University of Michigan and, at the time, the expectation was that the top-performing students automatically went into internal medicine. That was considered the “prestige track,” but I wasn’t very inspired by it. A friend of mine asked, “Have you considered pathology?” I hadn’t given it a moment’s thought. My friend said, “Pathology is great. You don’t have to pick a particular specialty; you can do whatever you want. It’s a research-oriented profession, so you can have a full-time lab and spend 20 percent of your time performing your service duties on the autopsy service.” That seemed like a good idea to me, so I switched my path to pathology, and it has served me well ever since.
I had some very good avenues into research early on. I had the great fortune to work in Judah Folkman’s lab – and he, of course, was the founder of angiogenesis research, so that was an amazing experience. After that, I had another fantastic opportunity while at Michigan to work for John Niederhuber, who eventually became the Head of the National Cancer Institute. Eventually, I made my way to Duke University as an assistant professor, where I was able to establish my own cell biology laboratory while spending (less than) 20 percent of my time on the autopsy service – exactly what my friend had originally predicted!
Finding inspiration in imaging
Before arriving at Duke, I was a Wilmot Cancer Research Fellow at the University of Rochester, working in the laboratory of Donald Young. He had developed a technique called “giant two-dimensional gel electrophoresis.” Regular 2D gel electrophoresis was performed on postcard-sized gels – but his gels were so large that we had to use chest X-ray film to develop the autoradiographs! Analysis, especially quantitative analysis, was a challenge; there were two to three thousand grey-to-black spots (translated proteins) on each film, and in the mid-1980s, we really had no straightforward way to determine how dark each spot was. One day, it occurred to me that we could try to build something we subsequently called a “pen-sitometer.”
The idea was to put the autoradiograph on a support table with a little hole in it and a light source below the hole, and then take the device – which was originally shaped more or less like a pen with a photodiode at the tip – and rub it over each spot we were interested in and it would tell us how dark the spot was. Don Young figured out that you could actually use a VIC-20, the predecessor to the only slightly less ancient Commodore 64 personal computer and, by hooking up the pen-sitometer directly into the game-port analog-to-digital converter, send a digital stream directly into the PC. We programmed the whole thing in Commodore BASIC – another ancient relic, but still the peak of my coding experience. Our setup allowed us to do some really serious research, though – for instance, we quantitated proteins that were responding to growth factors and steroids in cells in tissue culture.
That was my first introduction to the problems of image analysis. How do you capture information in image form and then extract useful data from it? That’s why, when I left Don’s lab and went to Duke, I took an adjunct appointment in the computer science department – so I could continue to work on image analysis.
Quite early on, I was interested in the use of confocal microscopy as a tool in pathology. In fact, I was the proud middle author of a paper on the subject back in the early 1990s. That was the genesis of my interest in optics, and my inspiration to move into technology development. Truth be told, my heart was in the tools, although I only realized that late in life. My next destination was Carnegie Mellon University, which doesn’t have a medical school. There, I worked on a technique called multispectral imaging, which is now part of the armamentarium for people trying to do multiplexed immunohistochemistry or immunofluorescence for cancer immunotherapy. And after that, I made the transition into industry, spending 10 years at Cambridge Research and Instrumentation (recently spun out of PerkinElmer). I was involved in developing the leading multispectral whole-slide scanner for multiplexed imaging, whose descendants are still commercially available.
After my stint at CRI, I consulted for 3 years. One day I got a phone call: “Would I like to be a professor at UC Davis?” My future chair, Dr. Lydia Howell, had a vision of bringing an emphasis on novel, imaging-based technologies into pathology, and fortunately I had come to her attention. I have had the distinct pleasure and opportunity to work with her and colleagues at UC Davis Health for the last seven years.
Currently, I work on microscopy with ultraviolet (UV) surface excitation, or MUSE microscopy, which has turned out to be an interesting and powerful new way of looking at tissues. Just after I arrived at Davis, a friend and colleague of mine from Lawrence Livermore National Laboratory showed me the work that he had been doing on UV-based imaging of tissues and pointed out the basic principle of it: namely, that ultraviolet light at the right wavelengths only penetrates tissue a few microns deep. That allows you to take a big chunk of tissue and image just a thin section of it from the surface down – approximately the same depth as a microscope slide.
Of course, I thought it was great – but he was mostly doing in vivo imaging via autofluorescence, to which I said, “I’m a pathologist. I can cheat and use stains.” Together, we started trialing substances that stain tissue in more or less the same way as hematoxylin and eosin, but are fluorescent. The result? It turns out that MUSE allows you to take almost any piece of tissue, cut a flat surface with a scalpel or razor blade, and still generate microscopy results that look as good as – or even better than – an H&E slide… in under three minutes.
I’m looking forward to seeing how MUSE microscopy plays out, because it solves a lot of logistical problems in pathology – speed, cost, and, most importantly, availability of histology facilities. And it’s not the only technique my colleagues and I are working on at the moment. Others – although it’s too soon to talk about them – are based on new ways of extracting additional information from existing slides. All I can say about that is, “Stay tuned!”
Digital pathology: proceed with caution
For now, I would have to say that the prospects are guarded. Why? For two reasons.
1. The business case
There are great advantages to going digital, most of which are logistical – not having to track down missing slides or repair broken ones, for instance, or having an easy way to share information across long distances. But the sad fact is that, currently, switching to digital involves a very large capital outlay, lots of retraining, and ongoing expenses, such as equipment maintenance, data storage, and so on. It’s hard to come up with a realistic return on investment, depending on your financial environment and how costs are calculated and allocated.
My own university has no immediate plans to go digital, and I think the institutions that have are still relatively few and far between. That’s because our transition is not like radiology’s. When radiology went digital, it replaced procedures; it replaced film. But when pathology went digital, it needed additional equipment and handling steps. Instead of eliminating slide preparation, the digital transition added another level of complexity to scanning, viewing, and storing information. So it’s not a simple story – especially not when trying to convince those who hold the purse strings. There needs to be a solid story on how money is saved by increasing pathologist efficiency and eliminating the problems of finding, storing, and retrieving slides – but it’s by no means a slam-dunk.
Digitization feels modern. It feels technical. It feels like a good solution. But as a way of serving up images to pathologists to make manual diagnoses, it has its costs as well as its benefits, and every pathologist and institution has to weigh that up. Jennifer Hunt, chair of pathology at the University of Arkansas, said at one meeting, “Digital pathology is not going to take off until you can get rid of the slide.” Perhaps one day, technologies like MUSE microscopy will help us reach that point.
2. The nascent technology
Digital pathology provides transformational value when you add the computer and automated or enhanced analysis. Essentially, that means it’s valuable when the computer can do things the pathologist cannot do alone or do efficiently. Some of those applications already exist; for instance, we have computational support for quantitating nuclear staining (ER/PR, Ki-67, and so on). These are not things that humans do particularly well, but computers – if properly programmed and utilized – do. Unfortunately, these applications are seen more as adjunct tools; on their own, they don’t make a billion-dollar industry
So will artificial intelligence (AI) sweep in and provide value when our slides are digital? Eventually, I anticipate that we will see automated diagnostics and enhanced prognoses, because computational tools can “see” patterns that humans can’t. We humans also have a limited professional lifespan; we only see so many slides and cases over the course of a career, so we may be perplexed when we see something we haven’t previously encountered. Computers don’t have that problem; properly trained, they can be exposed to everything we, as a collective, know about, so they can be – at least theoretically – more expert than the best human.
But AI is still fragile. Because there are so many different use cases, each one is a small application that takes a great deal of work to actually bring to clinical utility and achieve regulatory approval. AI with subspecialty expertise would be expensive to develop and validate across multiple institutions. With such a vast range of different computers, laboratory information systems, images, formats, and reports, it’s hard at this point in the technology’s evolution to imagine it playing out in real life, as opposed to in the academic laboratory. That said, once the FDA gives its blessings to the first anatomic pathology application, the equation will begin to change.
Why do I care about statistics?
This is going to sound very old-school, but statistics should really undergird what we do. In other words, we should have a reasonably refined understanding of what it means to obtain a particular test result in a particular situation, and of how to interpret that in the real world, where things like prior odds and posterior odds really affect the meaning of a test.
Pathologists are responsible for assembling a patient’s clinical and histological data and presenting their conclusions to the clinicians, who then move forward with treatment. Unfortunately, it’s often clear that humans don’t necessarily know how to combine information elements properly. Doctors may chase single aberrant lab tests or have difficulty integrating potentially contradictory data from multiple sources (DNA, RNA, histology, immunohistochemistry, and any other lab tests, plus the clinical situation).
That’s why it’s so important to understand that the world is a statistical environment, and that our intuition is usually wrong because we don’t understand probabilities and risks and benefits. In fact, presentation matters as much as substance. If you present data in a certain way, it suggests a corresponding response – but if you then take the same data and express it in a different way, you get a different response. “70 percent of patients benefit from this intervention” sounds very different to “30 percent of patients received no benefits.” Thanks to human psychology, expressing exactly the same information in different ways can lead to different outcomes. And yet, there is almost no training in probability statistics in medical school, or even in university.
Roll with the punches
People sometimes ask how I maintain a work/life balance. Truthfully, I’ve found that work is what I like to do best. Email regrettably fills many available hours and, when I’m not doing that, I’m catching up on reading (short) articles that keep me up to date without being overwhelmed. Also, interacting with other researchers and pathologists at conferences (local, national, and international) is tremendously fun and inspiring. I enjoy going from project to project, always looking for new experiences at and beyond the limits of my knowledge. Doing new things, especially when I lack the relevant credentials, allows me to collaborate and learn from others. My wife (and cats) are long-suffering, but we (minus the cats) do manage to get up to the mountains or off to Ireland or Australia when possible.
If I could go back to the start of my career and give myself some advice, I would say three things. One: “Go to class.” I slept through most of my classes. One semester, I attended three lectures from an entire physics course. Two: “Don’t plan.” Things will happen and you will adapt – because where I am now was certainly not planned. And three: “Don’t worry if you don’t know something—learn the vocabulary and collaborate.”
Professor and Vice Chair for Strategic Technologies in the Department of Pathology and Laboratory Medicine at the University of California Davis Health, Sacramento, USA.