Applying Artificial Intelligence (AI) to PSMA-targeted PET Imaging Agent 18F-DCFPyL in Prostate Cancer: Lecture & Discussion- Matthew Rettig
January 13, 2021
Applying Artificial Intelligence (AI) to PSMA-targeting PET Imaging Agent 18F-DCFPyL (PYL) in Prostate Cancer: Lecture & Discussion. - Matthew Rettig
Independent Medical Education Initiative Supported by Progenics Pharmaceuticals, Inc. a subsidiary of Lantheus Holdings, Inc.
Matthew Rettig, MD. is Professor in the Department of Medicine and Urology, and Chief of the Division of Hematology-Oncology, the Institute of Urologic Oncology, UCLA, Los Angeles, California. Dr. Rettig is a medical oncologist with an emphasis on the management of patients with advanced prostate cancer. He is the Medical Director, Prostate Cancer Program, David Geffen School of Medicine at UCLA, and the Co-Chair of the VA’s national Precision Oncology Program Cancer of the Prostate (POPCaP).
Neal Shore, MD, FACS, is the Medical Director of the Carolina Urologic Research Center. He practices with Atlantic Urology Clinics in Myrtle Beach, South Carolina
Phillip J. Koo, MD, FACS Division Chief of Diagnostic Imaging at the Banner MD Anderson Cancer Center in Arizona.
Alicia Morgans, MD, MPH Associate Professor of Medicine in the Division of Hematology/Oncology at the Northwestern University Feinberg School of Medicine in Chicago, Illinois.
View Complete Educational Program: A Step Towards Personalized Medicine: PET-PSMA Imaging in Prostate Cancer
Phillip Koo: Hello, I'm Phillip Koo from Banner MD Anderson, and welcome to another segment of our special program titled "A Step Towards Personalized Medicine, PET-PSMA Imaging in Prostate Cancer". Today we're talking about a topic that I think is very exciting throughout all of medicine, maybe more so in imaging, which is artificial intelligence. And it's exciting to know that there's being worked on in this space with regards to PyL PSMA PET imaging. And to speak about this topic, we're very fortunate to have with us today, Dr. Matthew Rettig, who is a Professor of Medicine in Neurology and also the Medical Director of the Prostate Cancer Program at UCLA. So welcome, Dr. Rettig.
Matthew Rettig: Thank you for the invitation to speak today. And I'll be talking about the application of artificial intelligence or AI to PSMA PET-CT imaging in prostate cancer. So here are my disclosures, and here's the outline. Just so everyone's on the same page, I'm going to provide some background about the performance characteristics of PSMA scanning as compared to other novel PET scans, as well as conventional imaging. And then I'll discuss a project that we're working on in which we're applying artificial intelligence to an automated interpretation of the PSMA scan. So PSMA scans have a consensus imaging approach, which is called the prostate cancer molecular imaging standardized evaluation, or PROMISE. And when we automate it with artificial intelligence, we call it aPROMISE for automated. And then I'll briefly touch upon some work that we're doing with AI to predict the onset of metastatic disease in patients with high-risk localized prostate cancer based purely on the image of the primary tumor.
So there are a number of PSMA PET probes that are available today. I generally think of all of these as similar in their effectiveness for imaging. So for the purpose of this talk, when we talk about PSMA probes, really just talking about them in general, as it applies to any one of these imaging agents. The two that are commonly used are the gallium PSMA-11, which just received FDA approval. And then the F18 PyL probe from Progenics. And this is one that we've used frequently, given that Progenics, the manufacturer of this probe, has a program to supply the probe for free to the academic community.
So PSMA PET scans are probably the most efficient and advanced PET imaging for detecting prostate cancer. It's superior to conventional imaging and on the right, you'll see the various references that support this concept as well as superior to choline PET and fluciclovine or Axumin® PET. So in the order of sensitivity, I find that PSMA PET is the best followed by fluciclovine, PET choline PET, and then conventional CT and bone scan or MRI and bone scan.
So this is an example of our experience at UCLA using gallium-68 PSMA PET in patients with intermediate or high-risk prostate cancer. So what you can see is that for the majority of cases, there's concordance between the conventional imaging and the PSMA PET CT scan. There's non-concordance in about 30% of the patients with most of the non-concordant cases involving upstaging from the conventional imaging to PSMA imaging. There's a smaller percentage of patients who are downstaged, including downstaging of metastatic disease to non-metastatic disease, which would obviously have a major impact on management and also prognosis as well as the patient's psychological wellbeing. When they're told they initially have metastatic disease, and then you can tell them they have non-metastatic disease. That's clearly a change in their prognosis and a very wonderful thing that you can get information that you can give to a patient. I've done that on a number of occasions, and it's really a transformative piece of information that you can tell a patient you have non-metastatic prostate cancer when they originally told they had metastatic disease.
So this is some work that we did at the VA. This is using the F18 PyL probe from Progenics. And basically what we're showing here is, on the left, is our pre-imaging staging, and this just sort of standard staging on the left. And on the right would be, on top rather, is the post-imaging, post PSMA imaging. And without going into too much of the details, what we see is that about 35% of the patients have either upstaging or downstaging with most of the patients about 30% having upstaging and about 5% having downstaging. And this actually changed the treatment plan and the vast majority of these patients, and that could be changing the radiation field, the decision to perform pelvic lymphadenectomy, or downstaging the patient and therefore utilizing a curative intent therapy as opposed to systemic therapy alone for palliation.
In a randomized controlled trial, similar data have been generated. And essentially in red, you can see on the right, the improved sensitivity of PSMA imaging for both pelvic nodal disease, as well as distant metastatic disease. And at the bottom, you can see the overall data. So the PSMA scan has a much higher sensitivity than conventional imaging and as well as a greater specificity and overall accuracy. And gallium PSMA-11 was just FDA approved. It was submitted as an NDA earlier this year by UCLA and UCSF based upon our experience. And these are the indications. So the gallium PSMA-11 is indicated for patients with suspected metastasis who are candidates for initial definitive therapy, so initial staging, if you will, as well as for patients who have a recurrence, a biochemical recurrence based on elevated PSA. And in these contexts, the PSMA scan has very robust performance characteristics.
Just to give you a sense of the report performance characteristics for patients who have a biochemical recurrence, about 50% of patients will have a lesion detected post-radical prostatectomy when the PSA is between 0.2 and 1. So it's a sensitivity of about 50%. And above one, the sensitivity is 90% or greater with the sensitivity increasing with increasing PSA.
So let me turn now to the artificial intelligence and the first part of this discussion, we'll focus on the utilization of AI to perform automated PROMISE reads. So to do this, what we have to do is first teach the algorithm what each organ is, including all of the bones, as well as the soft tissue organs, including lymph nodes and major viscera. And this is of course critical because ultimately what we're trying to do is match the PSMA image with a given organ to stage the patient according to, for example, the TNM staging criteria.
So what you can see here is, going from left to right, as the model is trained with more and more scans, it becomes more and more robust in detecting in this case, the vertebral bodies, and each color represents a different bone. And ultimately this was done on hundreds of scans, including scans that isolated specific organs. And we can see that we can segment the various organs, 27 organs, and the bones to get this nice, colorful picture that allows us to contextualize the PSMA image with a specific organ or bone. What I think is really helpful with this automated segmentation process is that it allows us to isolate the prostate signal and distinguish it from the bladder. So essentially one can subtract the bladder image and get the prostate image, which is, I think, very critical because a lot of the probe will accumulate in the bladder, which can make reading the prostate a challenge. And this is just an example of a patient who underwent the PSMA imaging and the automated PROMISE. And you can see the prostate here at the bottom, looking at the right here, as well as lymph node involvement. So very nice clear picture of the segmented organs.
So we undertook an initial hypothesis-generating study to understand the performance characteristics of the automated PROMISE, the aPROMISE PSMA read, and compare it to conventional imaging. And we utilized two independent readers for this. We used one novice inexperienced reader and one experienced reader of PSMA scans. And what you see here is the reading from the first reader, the novice reader, the inexperienced reader. And you can see that although local imaging called about 20 patients as non-metastatic without regional nodes, 14 of those patients were upstaged to nodal or distant metastatic disease. In addition, there were four patients who had metastatic disease, either lymph nodes 1A a or bone involvement in 1B, who were downstaged on PSMA automated PROMISE interpretation. Of course, these are going to be patients who would have changes in management based upon the change in the interpretation of the staging.
And this is the experienced reader, and you can see that there are similar numbers here of upstaging and downstaging suggesting that the inexperienced reader and the experienced reader are seeing similar images and calling them in a similar fashion. And when we compared the two directly, which is shown here, reader one versus reader two, if you look on the diagonal here, you can see that most of the reads are concordant between the two readers. There are some reads that are upstaged from the inexperienced reader to the experienced reader, shown here on the left column. And conversely, there are some that are downstaged from the inexperienced reader to the experienced reader, and this concordance is similar or better than the concordance that we observe with conventional imaging and published reports.
And importantly, the actual duration of time it takes to do an aPROMISE read is quite short. It's only about three minutes, three and a half minutes for each reader. Of course, there's a range depending upon the complexity of the scan. And essentially what happens with the scan is that the algorithm identifies the different lesions and it allows the reader to ask the question, is that real or not? So it's not completely automated, it's more of an assisted read. The readers have the ability to make a judgment call as to whether or not they think a particular organ is involved. And that ultimately leads to the final aPROMISE read.
So I'm not of course going to go into the details of these equations, but the point of it is that we can quantitate the PSMA reads. So we know what the SUV signal is in each lesion. We know the SUV max, and we know the volume of the lesion, and we can use those two figures together, the intensity and the volume, to come up with a PSMA index and some, the PSMA indices of all the different lesions to come up with an overall PSMA index. And this could be a potentially useful way of monitoring patients on therapy. So trying to determine responses to treatment can be a challenge, especially in bones. And this kind of approach may allow us to make quantitative assessments of changes in the PSMA image over time, including in response to therapy. This is yet to be tested prospectively.
So I'm going to change gears here and discuss our deep learning algorithm to predict coexisting metastatic disease using the intro prostatic PSMA image only. This was done in veterans with prostate cancer. So the ground proof from metastatic disease was conventional imaging. And what we did is we trained the computer to correlate the PSMA image in the prostate to whether or not the patient had concurrent metastatic disease. What we found was one, this is feasible. And two is that it has a high predictive capability. So if we just look at the SUV intensity, the intensity of the primary image alone, it doesn't work. Essentially with an AUC of 0.53, this is like a flip of the coin to determine or predict whether or not the patient has metastatic disease.
But when we use the AI algorithm, we see that the predictive power goes up significantly with an AUC of 0.81, and this was a cross-validation within our dataset. But we also use an external dataset to further validate the AI algorithm and found a very similar AUC of 0.82. So of course, this is a process where we're training the algorithm to detect concurrent metastatic disease. What we really want to do is have the algorithm look at the primary tumor in patients with localized disease to ultimately predict the onset of metastatic disease. And we think that we can do this using this algorithm that is further trained on our appropriate cohort of patients.
With respect to that cohort, we have these patients. We have patients who have localized prostate cancer who underwent PSMA imaging at the time of diagnosis with up to five years of follow-up. Of course, we need a lot of followup for these patients because the percentage of patients who have a recurrence, metastatic recurrence by say five years is only about 5 to 10%, but we have enough patients to get started to train this program. And this could potentially be a very useful tool if it works out to stratify patients according to risk for metastatic progression. And I should point out that when we utilize standard clinical characteristics, the PSMA imaging algorithm of the primary outperformed clinical criteria, such as a clinical-stage grade PSA level in predicting the onset of metastatic disease.
So this is leading to our prospective study that I just mentioned in which we're using the PyL, the F18 PET imaging deep learning model to predict metastatic progression after curative intent therapy in patients with a high risk of prostate cancer. And we're going to compare its efficacy alone, by itself compared to standard clinical pathologic features, as well as some of the predictive genomic models that are commercially available. And we'll of course also combine these different models to see if we can improve on the overall predictive characteristics and performance of this algorithm. So I'd like to stop there and acknowledge all of my collaborators at UCLA and the VA. I'd like to point out that Nick Nickols and Jeremie Calais, as well as Johannes Czernin, were critical collaborators on this project. And then my collaboration at EXINI, which was formerly part of Progenics, where I worked with Aseem Anand and Kerstin Johnsson in generating the AI algorithm. So thank you for your attention and I'm happy to answer any questions.
Phillip Koo: Great, thank you, Dr. Rettig, for that fascinating discussion about AI and PSMA PET imaging.
Phillip Koo: Today we're very fortunate to have with us Doctors Neal Shore and Alicia Morgans to give a 360-degree perspective on AI in prostate cancer. So when we talk about AI, I think from a radiologist's perspective, it initially created a lot of fear in our specialty, and rightfully so, but I think what we're learning today is the fact that AI can be an assistive tool. And I think Dr. Rettig pointed out how this could be a real powerful ancillary device, a supportive device, to help improve the efficiency and accuracy for a practicing radiologist or nuclear medicine physician.
Number two, I think if it improves patient care, it's really a no-brainer. We have to do it. And then finally, there's no question, all of us need to be part of that whole development process to make sure we can create the appropriate products that work best for our patients. And I think that's something that we encourage everyone to be involved with just because it is going to be a very powerful tool. So Alicia, from your perspective, what here excites you or resonates with you with regards to what the future state of AI might look like from a medical oncologist perspective?
Alicia Morgans: Great. Well, thanks, Phil. So first I want to say, of course, thank you so much Dr. Rettig for such a really enlightening conversation. I think that we in medical oncology don't hear enough about how AI might be used as this supplemental tool to assist our nuclear medicine colleagues as we continue to move forward. And I also want to congratulate you for doing so much work within the VA system, which we know it's a feat in itself to sometimes get some of that work done within that system. But it's such a great opportunity to work with these guys and gals, but gals, of course, don't have prostate cancer, but to work with these veterans to really do what we can to include them in these studies.
So one of the things that I find most exciting related to AI is really this thought that we might be able to better prognosticate which patients are going to be those that develop metastatic disease. And I love the work that you're doing in that area and really refining that algorithm. And I also think it's really interesting that you're going to then compare that algorithm against our currently used clinical tools which we've seen in multiple trials, at least adjuvant trials, where we're trying to intensify therapy that these clinical risk factors are only sort of reasonably good at best. I wonder, are you thinking at all of combining your work with molecular classifiers and things that are trying to kind of take that risk prediction to another level because maybe the combination could be more powerful?
Matthew Rettig: Yeah. I was actually going to include that in my talk. It's a great question. And yes, we are doing that. So to be specific, what we're trying to do is allow the AI algorithm to learn what the underlying genomic lesions are. So what we have is a fairly large cohort of patients who've had metastatic disease and have also had a metastatic biopsy. So we can teach the algorithm what the PSMA image looks like in a patient, for example, who has a homologous recombination deficiency, such as BRCA2 mutations, so that the algorithm is able to access all kinds of information that are not accessible to the human brain. So we hope to be able to do this. We have a lot of patients already. Of course, we need some more to increase the robustness of this program, so we're going to be doing prospective biopsies and PSMA imaging on our cohort as well.
Alicia Morgans: Fascinating to think that a BRCA-positive patient may have a different imaging output. You're just blowing my mind. So thank you, Dr. Rettig. Keep up the good work.
Phillip Koo: And when we think about imaging, I actually think there's a lot more in the images that we haven't tapped into. Oftentimes, there are multiple parameters. There's SUV max. There's SUV mean. There's texture analysis. There are so many different parameters but we tend to just look at it in size or SUV max, which to me, I think SUV max is probably the worst single tool we have because it literally just takes one pixel out of thousands of pixels and assigns a value. So I think there's a lot of work we can do and it'll be interesting when those two worlds meet and we use computers that obviously have more capability than we do to sort of put all this together. So Neal, from your perspective, urology, how is AI being discussed in the urology communities and where do you think we should be heading moving forward?
Neal Shore: Yeah. Thank you, Phil. And I agree with Alicia. Matt, great talk. Really, this is very disruptive technology and I think that's fantastic because artificial intelligence, it's here. It's not going away. And I think to your point, Phil, our nuclear medicine radiology colleagues really want to embrace this. Your question regarding the breadth of AI throughout urology, we're seeing it in many applications in histopathology and how we see a similar interpreter variability amongst pathologists, amongst radiologists. We see this with reads historically, even in very specialized labs. We see this with experts in the field in GU pathology. We also see the same discordance of interpreter variability with MRI. And this has been a persistent issue.
I think bringing aPROMISE and the work that Matt Rettig and others are doing changes the concept in... I say this with all due respect. We'll say, well, the department of "unclear" medicine, nuclear medicine. But frankly, nuclear medicine is moving way beyond this vagary of uncertainty, and nuclear medicine is really becoming a pivotal specialty now for us to understand, "Does the patient have metastatic disease that we traditionally were poorly measuring?". Clearly, your UCLA veteran's data is showing us this, and there are about a 30% discordance, and three-quarters of the time, it's an upgrading. A quarter of the time, it's a downgrading. These are really important issues when it comes to patient care. So I think AI, for all aspects of oncology, uro-oncology clearly, must be embraced from a pathology standpoint and a radiology standpoint.
My question for you, Matt, is... Blue sky thinking about this, getting the Grade A, level I evidence at some point in time, which it seems highly likely to achieve. You've already pointed out that it could be transformative regarding the way we use our genomic molecular classifiers. It might be additive. It might be superior. It might create a value proposition. But the other thing I just really wanted to ask you is, given this transformative concept, and you may want to answer this also, Phil, is how will this be rolled out and introduced in a broader sense to the nuclear medicine radiology community?
Matthew Rettig: Yeah. So the first thing that we have to do is validate these AI algorithms in prospective studies, and we're trying to do that now. We're starting to piggyback this aPROMISE onto a Phase III study that is based on PSMA targeting for the treatment of prostate cancer. So we're trying to validate this in a Phase III study. And ultimately once it is validated, assuming that it is validated, then the rolling out is a challenge. It requires education of its very existence, and then it requires the community to accept it. The actual utilization of the AI is not that challenging. The program is made to work very easily with the user. The interface is user-friendly. So I think that the biggest challenge is getting acceptance across the nuclear medicine and radiology communities.
One of the keys, I think, is to emphasize that this is not replacing the reader. It's assisting the reader. And in fact, if you look at this at least from the commercial standpoint, the financial standpoint, it may enhance financial aspects of nuclear medicine radiology by reducing the time to complete a read. We saw that the reads here are just a few minutes to do a read on a metastatic PSMA image, especially since it's pre-identifying all of the lesions and matching them with a specific organ. So I think it's going to take time and like any new technology, any disruptive technology, it takes time for the world to accept it.
Phillip Koo: Yeah. And I agree with Dr. Rettig. I think the acceptance piece is most challenging, and I think we do need to figure out what the financial models are going to look like around AI. I think the financial models, as Dr. Rettig mentioned, will likely be different if we're looking at a fee-for-service model versus a value-based model. But overall, I think the efficiency, it does improve.
That being said, I think this is going to push nuclear medicine physicians and radiologists to actually change their practice. Our skills should not be in the ability to detect something on a screen. Our skill should be in understanding the entire disease state and working closely with the med oncs, rad oncs, and surg oncs to really understand what the imaging means. And I think oftentimes, there's a gap today with regards to what we do with the imaging and how we interpret it and how we implement or take action on those results. And that to me, I think, is going to afford imaging specialists more time to start getting into this. And that's why I think down the road, even in nuclear medicine, we need to subspecialize within these imaging specialties even more so we could add more value to our colleagues. So I'll turn it around to the panelists for any final, last words, comments or questions. So I'll start with you, Alicia.
Alicia Morgans: Sure. Well, Matt, I'm just curious. One of the things that we see on these scans, of course, is PSMA-positive lesions, but sometimes there's PSMA-negative disease. And I really rely on my nuc med docs to tell me both what's lighting up and what then anatomically is also abnormal. And so do the algorithms reflect on any PSMA-negative disease? And just to comment to Phil, I love your idea or your thought that as integral members of the team, the nuclear medicine physicians help us understand the entire picture and part of that is this PSMA-negative disease. So what do you think, Matt?
Matthew Rettig: Yeah. So the algorithm does not allow for the identification of PSMA-negative lesions on the conventional or CTE component of the PSMA PET scan. In principle, we could combine approaches. For example, of course, this is getting expensive but if you do a glucose PET concurrently with a PSMA PET, there is evidence that the PSMA-negative lesions may actually be more hyperproliferative and hypermetabolic and take up the glucose as opposed to the PSMA. So this is, I think, an important issue, especially as we get to more advanced metastatic disease where there can be more and more disease heterogeneity. So that's an issue that needs to be addressed moving forward.
Alicia Morgans: Great. Thank you.
Phillip Koo: Dr. Shore?
Neal Shore: You reviewed very nicely, Matt, the recent remarkable and a very important approval by the FDA, the NDA that was jointly submitted by UCLA and UCSF, which is a real testimony to their leadership. I'm sure that the use of PET PSMA, which has been probably remarkably heavy at both of these academic institutions, will become even more so now given this recent approval, and the approval for both high-risk concerns of metastasis for localized disease, as well as for BCR, is a very generous approval. What can you tell our colleagues across the country, how this approval... As it relates specifically to UCLA and UCF, how that might help gain greater access in the early part of 2021 for our other colleagues who might want to avail themselves to this and get PET PSMA when they can't get a patient to travel to California?
Matthew Rettig: Right. Yeah. This is a critical question is, of course, how do we disseminate this technology so that everyone has access to it. I don't have a great answer for that. Even at UCLA and UCSF, there's a lag between the FDA approval and insurance coverage, including Medicare coverage for it. Ultimately, this is a regulatory question that I myself have. I don't have a good answer for that, Neal. And I think it's of course critical so that the access is not limited just to two institutions in California.
Phillip Koo: Great. So Dr. Rettig, any final comments or words to close this out?
Matthew Rettig: Yeah. So first of all, thank you for allowing me to speak on this topic. Spreading the word is I think critical ultimately for the acceptance of this technology by the community on a wide level. I think this technology offers promise, no pun intended. And I don't think it's any technology, any AI technology that we're going to implement, whether it's for imaging or pathology, is going to replace the respect of specialists. I think it's going to assist them. It's going to improve efficiency, and ultimately I think it will improve patient care. So I thank you again for allowing me to speak today.
Phillip Koo: Thank you very much for joining us.
Alicia Morgans: Thank you.