Artificial Intelligence In Digital Pathology

Welcome to our seventh episode of the Ask a Pathologist Podcast, Innovators in Action. We had the privilege to interview Dr. Todd Randolph about artificial intelligence and its current and future role in digital pathology.

Dr. Todd Randolph holds board certification in anatomic pathology and cytopathology. He has worked in pathology for nearly three decades now, and has been practicing digital pathology since 2016. Throughout his career, he has held pivotal roles in leading and overseeing laboratories within prestigious hospital systems and commercial laboratories across the United States. His expertise extends to clinical research and the rigorous validation of emerging diagnostic medical technologies. With a keen eye for bridging clinical laboratory applications and commercial imperatives, Dr. Randolph has accumulated extensive experience in achieving organizational objectives.

Submit questions for our future recordings here: https://lumeadigital.com/ask-a-pathologist-podcast/

Sponsored by the Digital Diagnostic Summit and Lumea.

Transcript

Jake Brown: Hello, and welcome to the Ask a Pathologist podcast. This is the Innovators in Action Edition, where we bring in guests who are moving the needle from theory to actual, implemented, and viable digital pathology. As we know, digital pathology has been around for quite some time, but we're interested in talking to people that have made it happen in their day-to-day work, and we're excited to have Dr. Randolph Todd Randolph on the podcast with us today. My name is Jake Brown. I'll be the host, and we're excited to have you, Dr. Randolph. Thanks for joining.

Dr. Todd Randolph: Thanks, Jake. It's my pleasure. We go way back.

Jake Brown: We do go way back. Hey Dr. Randolph, do you mind giving us a little brief overview of your experience and kind of where you've been, and where you are now?

Dr. Todd Randolph: Yeah, absolutely.

So I practiced, I guess now you would call it conventional, non-digital pathology, from 1996 to about 2016. Initially, in '96 it was a smaller practice, but it grew to be about 20 pathologists in the urban central region of the Salt Lake Valley.

And I practiced almost exclusively anatomic pathology, surgical pathology, cytopathology, a little bit of clinical pathology in that era, but I'm certified in AP/CP.

And then I did, we self-specialized in the group, so I started developing an interest in GU pathology in the late 90s, early 2000s. But then I had an opportunity. I kind of had done that for close to 20 years, conventional pathology, and had an opportunity to leap to a different career, different career path. So at that time, I joined a biotech company, on kind of a part-time basis, and then also began my career path in digital pathology. So that was about 2016. 

In the very early days of Lumea, and I knew the founder, Matt Leavitt, so when I made the transition away from my regular practice, my conventional practice, I started working with the digital platform back in 2016, and then that has expanded over the years subsequently in 2016. So I was active with the biotech company and digital pathology, kind of simultaneously for a number of years, but then in 2020, the spring of 2020, I made the transition to do strictly, well, almost exclusively, digital pathology. 

In that realm, I also was doing some GI pathology in a conventional format for a few years, but that subsequently is now entirely digital as well. So, I still do a little bit of consulting for the clinical biotech company, but almost exclusively digital pathology practice, which has allowed me to obviously be flexible where I practice and where I live. 

So, and then that's been advantageous. I work closely with another individual who is exclusively digital as well. That's primarily on the GU practice. So I do have kind of a comrade and a group of individuals who are also on the digital platform who can interact with me virtually looking at cases, sharing cases, getting confirmations, that type of deal, but essentially that, I was probably a pretty early adopter just because of where I was in 2016 and knowing Dr. Leavitt and working with Lumea, so it was an opportunity that was probably somewhat unique to me given my situation and circumstances. So, in a nutshell, that's kind of where I am now, yeah.

Jake Brown: Awesome. Well, that's awesome. 2016, I know digital pathology has been around longer than that, but I think you're going to be hard-pressed to find somebody that was clinically day-to-day using digital pathology in 2016. There's not many out there.

Dr. Todd Randolph: Probably not, yeah, exactly. Like I said, I was just lucky to be in the right place at the right time to adopt that technology.

You see it grow, see it mature to where it is now. 

Jake Brown: Yeah, yeah.

Dr. Todd Randolph: Where it was, yeah, for sure, yeah.

Jake Brown: Well, thank you for that background, and it's exciting to have you on. I mean, that much experience, man. We're grateful to have you. And the topic that we're excited to talk with you about today is actually on artificial intelligence. It's getting a lot of love in the news.

Dr. Todd Randolph: Absolutely.

Jake Brown: You can look up the Wall Street Journal and see $22 billion in revenue from one company last, you know, the fall of 2023. You can see successful implementations from like CHAT GPT and all these different things, you know, artificial intelligence is here to stay, not going anywhere. But what I find interesting, what we're excited to dive in with you and talk about, is artificial intelligence in the digital pathology space.

Specifically, you know, it seems we'll get your feedback on this, but it seems that the actual use day to day is actually quite minimal.

It doesn't seem like there's a huge, significant imprint in the digital pathology space day to day. And we're interested to see, you know, why is that? And then, also, will there be eventually, and what do you think about that? So, that's kind of what we'd like to talk about. But let's start with asking you, Dr. Randolph, why hasn't there been a tidal wave, if you would, with artificial intelligence in the pathology space as of today and right now?

Dr. Todd Randolph: Why has it been what, Jake, as far as a lack of enthusiasm, is that it?

Jake Brown: Yeah, lack of a tidal wave or like a lack of overwhelming adoption in day-to-day practice.

Dr. Todd Randolph: Yeah. Yeah. Yeah, I think it may be better if I just kind of go into my experience with it in that way. I think that will kind of answer my opinion of why it's not been adopted widely yet. 

So, as I mentioned, I've been practicing digitally for a while, and then I have experience with a couple algorithms mostly geared toward diagnosing GU specimens, almost exclusively prostates. So my experience using it, it was probably a solid six to 12 months, was that the algorithms, again, this is diagnostic algorithms. I'm gonna kind of diverge on that in a little bit. Yeah. 

My experience was is that they were overly sensitive when you look, at least, prostates. By that I mean they would indicate many areas on the prostate or biopsies that were benign, that were not, that the algorithm felt represented atypical cells, atypical areas, or potentially cancer. So from my perspective, that required me to spend additional time looking at higher resolution, higher magnification, more areas than I would just to be sure, oh, is it finding something that maybe I missed? And so, but it would also, you know, find small areas of cancers that may be a non-GU pathologist or somebody, you know, you would just miss, right? Single cells. So it was helpful in that regard, but the amount of time that you would have to spend to look at everything that it would indicate as being potentially worrisome, potentially malignant, would slow me down.

Um, so, so I think that that would be one aspect of why it's not been adopted, at least in particular for diagnostics. And that may have improved and it may be better for other tissue types. 

But my experience was with prostate, uh, and in that regard, so another reason is perhaps the expense, right? We were interested in negotiations with the company potentially to provide the support, although, you know, some of us used it, some of us didn't within the group setting, uh, but it was, it was almost cost prohibitive because I think the vendor felt that it was probably performing better than, than really what it was and that it would basically a full-time FTE pathologist or a half-time FTE pathologist, so it would be worth that much or close to that much in value.

So, um, so those I think we're probably the, at least, you know, that's where it was a year and a half ago. I've not done a lot with it since then. But that being said, I did find it comforting. This would be a reason why it would be used when it was entirely negative and I felt that the prostate biopsy was negative as well. It was reassuring for me to look at the case and then apply the algorithm and for it to say it doesn't see anything and I didn't see anything either. 

That doesn't necessarily answer your question of why it wouldn't be adopted. That would be a reason why it would be adopted. So, but that's kind of a, you know, we're talking about that particular application, the diagnostic application within this whole group of GU pathology, which is a high-value practice, you know, worldwide, prostates in particular. But I think that from my perspective, it would be cost and then who's going to pay that cost and then, you know, being overly sensitive.

Jake Brown: Yeah, yeah, that's that's really interesting. I hadn't thought about the oversensitivity factor, but when you describe that, that's just one of these things where day-to-day practice, that's going to add up when you're using it and you've got to wait.

You don't see anything, but it does, and then you've got to wait, zoom in, look through all those things, make sure that you're practicing pathology the way that you want it to be practiced, but it introduces an extra step, which I hadn't thought of. That's really fascinating. Obviously, those algorithms are always improving. That's one of the beauties of artificial intelligence, but something that I hadn't thought of, so interesting. 

And then cost, I thought of that one and heard, those are some hurdles that are still being kind of addressed in the market today. There's all these talks of CPT codes, reimbursements, stuff like that, that could offset some of those costs of using them, are probably in the future too wouldn't you say?

Dr. Todd Randolph: I think so. Yeah, I think so. I think… you know, depending upon what your comfort level is with various tissue types. In other words, if you have, you know, prostates is something that I look at a lot of. If you have a lot of experience and you're very comfortable with prostates, then, you know, the algorithm may have to be really, really sharp to improve on your performance as a pathologist. But if you don't see that many prostates, or that many breasts, or, you know, different types of tissues, it may be very helpful for you in the future as the algorithms improve to make the appropriate diagnosis to identify the very small lesions. So, I think ultimately, you can always use the term standard of care, right? 

If the algorithms improve to the point where it becomes standard of care, where the performance is proven to be much better than just the unassisted traditional pathologists' eye looking at the digital images, then I think there'll be greater reimbursement or reason for any institutions, in particular, to say, Hey, you know, we're using these advanced algorithms to make your diagnosis and finding lesions and making better diagnosis than, than in the hospital or the system across the street. 

So, so I think you have, you know from marketing saying we're using this because it's showing me better and then to you know helping pathologists with tissue types that they are not that helpful with you know. If you're a general pathologist out in the community or smaller community you may not see that many specimens. I mean for me that would be helpful. For example, like I said, I've specialized over the last almost, you know eight years now so I would maybe welcome having an AI algorithm for breast pathology for example finding small lesions. I haven't seen breasts, you know the high lines of breast for several years now. 

So I could see where if there's a good algorithm, a better algorithm, that I would use that and adopt that. Again, the cost would have to be reasonable, but from my perspective, you know I certainly could do that and the way that I've set up is that it would be a business expense for me as opposed to just, you know, out of the pocket type of, so it'd be pre-tax dollars as opposed to post-tax dollars. That might make sense for me to make that part of my practice in the future.

Jake Brown: Yeah, that's really interesting. I like what you had to say there about the standardization and being able to, as you implement those as standardization tools, and then as facilitators for, you know, pathologists that are less experienced in one thing, but are needing additional input and advice on something, you know, that they can have access to that with a click of a button instead of sending the slides out for, you know, a second opinion and waiting for a while to get something back. So, yeah, that makes sense, and those are really good thoughts.

You touched on something earlier that I want to come back to and you pretty early on in the conversation made a delineation between diagnostic AI and then kind of this bucket of would you say like other AI, more flow-driven as what I'm assuming you were going to say.

Dr. Todd Randolph: Yeah, exactly. Exactly. I was trying to, I've been thinking about this for several weeks and how do you characterize that and maybe I would call it a kind of a virtual assistant, a virtual pathology assistant and I would relate it, you know, in primary care pathology, or a primary care physician's approach would be, you walk into the office as a patient and be… whoever is, you know, the receptionist, whoever's there identifies you, looks at your license, says, okay, you know, you are who you say you are, right? So that's an identification process. And then, and then you walked through and you have somebody typically who takes your history, takes your blood pressure, and then maybe asks you a few questions about, you know, what's going on? Why are you here?

That would be probably, you know, a mid-level, like a physician's assistant type questions, right? Where they're coming up with a preliminary diagnosis, what they think is going on.

You know, earache, sore throat, pharyngitis, you know, whatever, they're coming up with this differential.

And then that person would then walk out to the physician, the most specialized, and give them a lowdown when maybe that physician is walking down the hallway, making incredible use of their time. 

So they're, you know, incredibly efficient. They're well prepared for when they knock on the door and enter and start to talk to the patient. 

They already know a lot about the patient. They may make some final observations, a couple of questions, agree, disagree with what was, you know, the preliminary diagnosis, and then off they go. So their time commitment, as opposed to, you know, the physician meeting the patient at the door. 

Saying oh, yeah, I missed this bit. I see your thing.

Let's go in and get your weight, let's get your blood pressure whatever that you know, the physician could do all that but that would not be an efficient use of their time. So in the pathology world, we really obviously don't have that direct patient interaction. But we are seeing patients via their samples every time we look at a case. So if you have, or you could trust an artificial intelligence to you know, especially now everything's digitized, right?

So in the conventional case, I would grab a slide and look at the label on the slide and say okay patient's name information matches the paperwork, the requisition or you know over the years it matches what's in the laboratory information system, but nonetheless, that's it. That's a check that I have to do and I'm still currently doing it with the digital, albeit faster right, the thumbnail at least the system that I use the thumbnail shows up, patient information also shows up in a separate category, but I still need to compare those two fields to make sure that they agree. I still most I do a lot of GI pathology. So I'm looking at the requisition. Why the patient's here, you know, is it a gastric pain. You know change and is there some other issue with the colon that's going on or whatever So I have to read that, or I should read that, so that takes me time to do that. So there's identification, there's why the patients here?

And then also, you know, if this laboratory information system's been around for a while connected, then there's previous biopsies, previous specimens on this patient that I can review and should be available. So those are three pieces of information that I look at fairly routinely on almost every case, right? But would there be a role for artificial intelligence where you would trust artificial intelligence? Well, I don't have to look at the thumbnail for the glass slide. It's already looked at. It's matched it. And it tells me if they don't match. But I assume it matches if it's showing me the case. So that'd be a time saver. 

Two, as I'm looking at the case, it could be like Tetchy Petit telling me the history. 

Right? I don't have to read it. It's to a microphone earphones or whatever. It tells me the next patient, that they are here for, whatever they covered here, you know, it doesn't have to be... they could just tell me what their symptoms were right away and then whether or not there's previous things to look at. And I could do a voice command and say okay. Let's look at the previous gastric biopsy and it just pops up right there for me. I don't have to make any clicks or whatever the AI just generates. Okay, here it is. Boom. So huge time saver there. 

So that if, and then if you put that together with the diagnosis then all of a sudden we have a patient presented to me, previous history whatever is going on, and then I potentially have as the algorithms improve prostates as I annotate them so it could be already annotated and I would just look at it say agree agree agree agree agree and then the report is already generated by the Artificial intelligence as well. A little more complicated with prostates than it would be for GI specimens - the GI specimens can be pretty straightforward, a lot of them.

So, if you could take that one step further, you could imagine that you have, you know, this is different than primary care, but you would have a whole series of images that I might look at. Not necessarily divided by patient, but just images. And I wouldn't have to necessarily look at patient identification on each of those images. 

And I could just run through those and say, yep, yep, yep, yep, you know, provided the diagnosis is there, tubular adenoma, you know, it's already there. It's not in the report or anything. It's just a tubular adenoma. And I say, yep, I agree. I agree. I agree. I agree. I agree. 

So, I could look at a large number of images very quickly without having to do all the administrative work. And then the AI would basically parse those out to those individual patients. So I could look at 40 or 50 images, which may be from 30 patients in a pretty quick amount of time if I'm not having to check all these things, put it all on the report, whatever. 

That would be huge. So essentially, what you've done is you've combined the diagnostic with the administrative assistant. 

So you almost have a virtual biology resident where you're working with that person. 

And a lot of times people will say, well, residents slow them down. And they can slow them down. But typically that's because they're not familiar with it or you're in a teaching institution and you're teaching the resident as you go along.

I don't worry about teaching artificial intelligence. That's not part of that role. 

So essentially, you're getting all the advantages of a resident, PA, you know, whatever, without having to do the instructional portion of it. 

So just looking at, you know, what's going to be available in artificial intelligence in the future, you know, it doesn't make sense to have an administrative assistant helping you through that as a pathologist, now it's too expensive, but if there's something can make me that efficient, and if I trust it, that to me is really the future - the application of the diagnosis plus, you know, the administrative component working together. And I don't know if any people really thought about the administrative component of artificial intelligence helping the pathologist, but that's, in each case, that's easily half of my time.

Jake Brown: Yeah.

Dr. Todd Randolph: Just confirming things, looking at things, generating the report, putting in the CPT codes, putting in the ICD-10 codes, and if that could all be done for me, and I just, like I said, the physician, the primary care is getting all this information while they're walking. It's already been collected. They can just look at it, click, click, click, click, okay, it looks good, boom, sounds good to me, off you go. That way you're, you know, I'm using the skills that I'm most trained for that the other people don't have, as opposed to duplicating the skills that they have.

So to me, and it's all, you know, it's artificial, it's quite artificial, but you know.

Jake Brown: That's how I know.

Dr. Todd Randolph: You know, I think initially you say, you know, can AI now, I don't know if it can the way it's configured now, but will it like I would could see where that, you know if you had the right company and as that technology becomes less expensive to apply you know because you know the requisitions are going to be in a different format for different practices, right? 

But you know, can AI scan through it and find okay where is that information if it's if you're in one practice and it's always the same that it can always look for the information in the same spot. This was always going to follow, you know patient history or patient information or whatever it's gonna it's just gonna do that. And it can do that, you know well before I look at the case. Where it can do that at night and then basically just condense it and tell me what I need to know in the following morning, and especially if I can interact with it, you know, as I'm working through a case, it's just like I'm talking to somebody across the microscope in the old days, and they can answer questions based on what I'm asking, and tell me, is it so? 

That's how I see it applied, where you have both diagnosis plus administrative assistance, which is what the pathologists do daily, right? They're making the diagnoses, but yet they're still going through all this information at the same time. And if you could improve both of those simultaneously, then you really have something, I think. And I'd be willing to pay for that because it would make me so much more efficient. 

Jake Brown: Yeah, I love this. I love this. Thank you so much for sharing that. My question As we wrap up the conversation today was gonna be in your ideal setting, You know like if you can just dream up the perfect scenario, What is it? But I don't need to ask that because you just - It makes a lot of sense to me and it's like, you know what Dr. Randolph you better, you know, you better get with the patent attorneys and

Dr. Todd Randolph: Hopefully somebody who listens has a start-up company in mind and then they can they can reach out.

Jake Brown: Yeah. Yeah. Hey, I mean about the experience and my thing that I love about this conversation today is the experience isn't that “oh, you've been in a larger institution where there's been a scanner and you've looked at a few cases for research here and there and sent them off to pharma for further development.” No, you've been diagnosing clinically and making a change in people's lives using a digital platform for close to eight years and I think that's extremely valuable and we're grateful for your time today and thought your insights and perspective has been amazing. Thank you.

Dr. Todd Randolph: Yeah, you're welcome.

Jake Brown: Yeah. Okay.

Any final parting words, Dr. Randolph, anything else you'd like to add to the conversation before we wrap up today?

Dr. Todd Randolph: I don't think so. You know, I think, you know, obviously, we're all endorsing digital pathology and I think it's certainly where it's going and I guess primarily because now we can actually use artificial intelligence to improve what we've done in the past, not only to make the pathologist more efficient but also offer improved diagnosis to patients. And then, you know, something we didn't get into was, you know, machine learning that, where you might be able to predict things based on images that are beyond what I can do as a pathologist currently, just because I have, you know, a limited number of neurons that are fading, you know, every day. But, you know, I think in the future, the artificial intelligence may see things that we can't see that will take, you know, regular light microscopy to a level that really hasn't been possible before the digital age.

I'm sure companies are working on that as well, but that's probably pretty cool stuff, too. 

Jake Brown: Yeah. Well, maybe we'll have to have you back on in a year and we'll readdress this and talk about all the cool new stuff that's come out and see what then what your future looks like. Eventually, it's going to be like, hey, we're living in that future.

Dr. Todd Randolph: That's good, Jake.

Jake Brown: All right. Thank you. 

Dr. Todd Randolph: You're welcome. 

Jake Brown: Have a great rest of your day. Thanks for joining us today.

Dr. Todd Randolph: That's good.

Tune in next month for our next podcast. Thank you to the sponsors of our program, Lumea and the Digital Diagnostic Summit, our listeners, and our guests for making this possible and for your support.