Have We Reached the Tipping Point for Digital Pathology? An Interview with Michael Bonham, MD, PhD, Chief Medical Officer of Proscia - Part 2 of 2

Dr. Michael Bonham, Chief Medical Officer of Proscia is back.  Last time we talked about how pathology has been slow to adopt new technology, using essentially the same equipment for the past one hundred fifty years.  Have we finally reached the tipping point, and is artificial intelligence the missing ingredient that will push us over the edge to full scale adoption of digital pathology?

…Continued from part 1 

Michael Bonham, MD, PhD: When you look at other lab tests we have, such as chemistries or some genomic tests, you're able to very precisely decide what's positive and what's negative.  This is something that is not well-suited for human eyes.  Image analysis has a lot of promise to improve the precision and accuracy and patient selection for companion diagnostic assays.  What we're seeing now is the ability to use more advanced algorithmic techniques, such as deep learning programs, which can improve upon existing traditional image analysis that has been on the market for the past ten or twenty years.

Joseph Anderson, MD: What does AI mean for the surgical pathologist?  Is it going to take their job away?  Is it going to make them more efficient?  Is it going to allow them to focus on more lofty tasks?

MB: I'll tell you how I think AI benefits the pathologist.  There are three primary areas: 1) efficiency gains, 2) standardizing the Pathology review and 3) providing precision and new insights to enhance patient care.

Efficiency, I think, is fairly easy for most people to understand.  It is, in some ways, automating some of the more routine tasks that pathologists have to do, such as searching for small areas of interest - the “needle in the haystack” problem, but also having an intelligent work flow. The same way that Pandora knows what song that you're going to like, if the AI is reading the cases before they come to the pathologist, they can be sorted in a way that's most preferable for the workflow.  There's a whole range of ways that that AI can be used in the workflow setting. We've seen this already being done in radiology.  The concepts are beginning to translate over into pathology. 

Standardization is another important area.  It gets back to being able to quantify what is on the slide - instead of having this subjectivity, where you say, “This is good, or this is bad.”  Let's take melanocytic lesions.  These are lesions that are on your skin.  Some of them are benign, like moles.  Some of them are malignant, like melanoma.  The spectrum between a mole and a melanoma is a gradient. Some biopsies are in between, and they're called “atypical.”  The problem is that if you ask two pathologists to look at the same biopsy, you'll see that the disagreement rates are often quite high – for many reasons.  But if you were able to run it through an AI algorithm and you have a score which told you something like, “This case is eighty percent likely to be melanoma,” for example, now you have an additional piece of data that you can use to help guide what your diagnosis is it going to be. AI doesn't give you a diagnosis, but it gives you additional information to help process the case.  The hope is that in the future, all pathologists will have the same data.  This will help to drive concordance of what is the actual true diagnosis.  That's how AI can help with standardization.  

Precision and new insights is a very exciting area, because it really gets into the fact that there's so much information lying within the slide.  The information in the images is really beyond the ability of a human being to calculate and correlate with outcome.  You can imagine that you could take the patterns and features in an image and correlate them with patient outcomes.  The key is that you have to have so many data points in order to make meaningful correlations.  Use of AI and computational pathology enables the possibility that through an image alone, we can have a greater sense and greater granularity in terms of what the prognosis may be what treatments may be appropriate for a particular patient.

JA: We spoke last time about how pathologists were very good at what they do and how they're very good at doing what’s been asked of them.  That's not to say, though, that there's not a tremendous amount of variability or ambiguity.  Melanocytic lesions, for example, is one area of pathology where some cases are challenging, and it may not be clear in some cases whether something is benign or malignant. In other areas, such as breast cancer and prostate cancer, making the diagnosis is not a problem, necessarily, but it may be a question of how aggressively the patient should be treated.  Is this going to be an indolent tumor or an aggressive tumor? Is this an area where we're going to be able to make a big improvement?

MB: Yes, exactly.  That's what I was getting at.  There are different challenges to various types of cancers.  Every disease has its own challenges.  We like to joke that there are twenty thousand problems to solve in pathology and each one can be handled separately.  Perhaps in one disease, the hard part is telling if something is benign or malignant.  In another disease, it may be easy to tell something like that, but it may be very hard to get good concordance among pathologists about the grade or some other features.  I believe that we will see some sort of prognosticators based on images.  It will be interesting to see how that compares to some of the prognostic testing we're seeing in genomics.  Are they complementary?  Are they synergistic?  Are they competitive in terms of utility?  We just don't know the answer yet, but I think the amount of data there suggests that we should be able to draw some correlations.

JA: We also spoke about regulatory barriers coming down.  There's been two FDA approvals in this country, I believe, for the primary diagnosis of tumors using digital pathology, and recently Proscia received a CE mark in Europe.  Can you tell us the significance of this?

MB: For Proscia, it means that we’re maturing as a company. We’ve developed an image quality management system. We are building software that is compliant with the stricter quality demands that are needed in healthcare.  It's an important milestone for us as a company to demonstrate that we can meet high standards. The CE mark, itself, shows that we have the quality controls, the process, and the documentation in place that allows us to market our software to help generate a primary diagnosis in areas of Europe and Latin America.  They will be using Concentric.  It is an important milestone for us.  We would like to cross other regulatory milestones as well - on both the primary diagnosis side as well as the AI application side.

JA: What would you say to pathologists or labs out there who are on the fence or considering going digital?  

MB: I think it's important for labs to really think about the future. You have to understand that digital pathology is not about trading the microscope for the monitor.  This isn't about hub-and-spoke models or saving money on FedEx packages.  Those things are nice, but they maybe squeeze out five or ten percent efficiency gains.  Digital pathology, in my mind, really needs to be thought of as a platform enabler for AI and all that it will bring.  

Digital pathology is the only way that labs will be able to fully participate in this technology revolution that is here now. We are building the solutions, and we want to help labs to take the next step.

JA: What challenges or hurdles you think we're going to face in the next decade?

MB: There are, in terms of using AI, some challenges. For the practicing pathologist who has been very good at what they do in terms of providing diagnostic information, they have very little technical understanding of the underlying algorithms that are being built using AI.  It's hard to understand.  If you don't understand how something works, it's hard to figure out what the results mean. As pathologists, part of our job is to assess new tests. Is this an accurate test? What is the error rate?  Is this a useful test?  In anatomic pathology, it’s hard to grasp what's happening with an algorithm.  That will be a difficulty in terms of bringing on AI clinically.  It just hasn't been done yet.  The field is sorting out what type of evidence is required - both in terms of primary risks and benefits.  How do we validate these AI applications in a clinical setting?  What is the ground truth that is used to generate these algorithms?  We haven’t settled on standards, but I will say that our field is working on it right now.

Those are probably the biggest barriers that I see.  There may be some reluctance over AI and new technologies being involved and, of course, fear of the unknown.  The more the field can provide information and data about how the algorithms work, and how they were generated, that can help us avoid the so-called “black box” problem.

JA: Mike Bonham Chief Medical Officer of Proscia, how can people learn more about you and Proscia? 

MB: I would encourage anyone who wants to learn more to go to Proscia.com . We also have a Twitter and Linkedin site.  You can also follow me on Linkedin if you're interested. I try to post useful articles and information.