false
Catalog
Prostate MRI and Molecular Imaging: Core and Advan ...
M1-CGU02-2022
M1-CGU02-2022
Back to course
[Please upgrade your browser to play this video content]
Video Transcription
Yeah, so I'm going to talk about PSMA-PET and PET-MR, and I'm actually gonna focus on the initial staging, because I thought for a prostate MRI session, I thought I'd focus on where PSMA-PET might be valuable. We already know that PSMA-PET is great for looking at metastatic disease, and that's the primary strength of PET. So the objectives, I'll look at advances in PET scanner and clinical indication, PET scanner and radiopharmaceutical development that has improved our ability in nuclear medicine to detect the extent of prostate cancer, clinical indications for PSMA-PET, and a few cases demonstrating how PET can actually complement MRI and potentially improve our ability to stage. So as you know, this is the complete list of imaging modalities that we use for prostate cancer, and the biggest splash now is, of course, PSMA-PET. I bring up the other PET agents because we actually did improve in our ability to identify small sites of prostate cancer, just not as well as we did with PSMA-PET. And what has contributed to our ability to detect smaller and smaller sites of disease? First, our PET scanner improvements, so digital time of flight, with improved sensitivity for detecting photons, statistically-based and other noise-reducing algorithms using AI are becoming more widely available on the market. The extended field of view PET-CTs are enabling us to see smaller and smaller sites, and this is particularly relevant for nodes, which is why I'm pointing this out. So the result is that we get high-resolution images that now we can characterize sub-centimeter lesions better. So in terms of radiopharmaceutical improvements, when you have a radiopharmaceutical that gives you higher signal-to-noise ratios, you know that's also going to improve your detection rate, which is why PSMA-PET has been so valuable. So this is a study looking at comparing fluciclovine, choline, and PSMA-PET. You can see PSMA was generally higher across all modalities, and we have several studies showing that PSMA-PET does improve our ability to stage patients. So the pro-PSMA trial showed that PSMA-PET was consistently better than the conventional bone scan and CT for initial staging and detecting full extent of disease. And then two trials looking at gallium-68, PSMA-11, and then the OSPRI trial next showed that we have very high specificity, positive and negative predictive value. The sensitivity for nodal metastases is still on the lower side, so on average 0.4. But when you look at node size that's greater than a centimeter, that sensitivity improves to about 0.6. This is, again, potentially using a mix of scanners, so this does not necessarily include the use of the most advanced PET scanners that we have available now. The OSPRI trial also was the primary data supporting the FDA approval for DCFPYL, which is Polarify. And so they also showed similar results with very high specificity, PPV and MPV. And then again, we had improved detection and sensitivity when you exclude nodes that are less than five millimeters. So that's an even lower threshold. Now we can detect nodes that are greater than half a centimeter, which is smaller than what we have usually, traditionally said with PET. We've used one centimeter, that's sort of been a practical cutoff for what we say we can reliably characterize with PET, but we are seeing smaller and smaller sites of disease. So just to show what improved radiopharmaceuticals can give you, these images were acquired on a non-time-of-flight PET-CT scanner. And you can see that there are small sites of uptake that are in pelvic nodes that are typical sites of metastases and then a left common iliac. So you can see this with some of the other, with those C11 choline and fluciclovine that we had before PSMA PET came along. But now with PSMA, because PSMA has higher signal-to-noise ratios, even compared to the other amino acid-based tracers, you can see very small lesions. So this is another maybe a 0.8 millimeter node that had increased uptake and was confirmed as a site of prostate cancer. And we can use PSMA PET to discriminate whether nodes are metastatic or not potentially. So with the caveats of those lower sensitivity thresholds. But when you see the uptake, that's why the specificity really has been a great value with PSMA PET. Because when you see that uptake, then there's a very high chance that that is a prostate cancer site. And so you can see that, and the distal external iliac nodes have always been a challenge with both C11 choline and fluciclovine. There tended to be nonspecific uptake, even if you saw asymmetries in that uptake. But PSMA PET tends to have more specificity than that. And so we can better assess those small nodes. So in terms of how it can complement MRI staging, I think you're gonna hear about all the advances in prostate MR. So obviously I don't have to go through that. And you do really well with prostate MRI in characterizing the full disease extent in the prostate for sure. So the question is, is PSMA PET required to look in the prostate? And we don't have a lot of strong data to say what the exact value is, but we definitely have cases throughout the literature showing that PSMA PET can sometimes complement what you see in MRI and vice versa. We are going to see PSMA negative disease or low PSMA expressing disease. And so that's where obviously PSMA PET is not gonna be as useful for those disease types. In this case, this is a patient who came in. The MRI had already been called this T2 hypo intense lesion with diffusion restriction. That showed very low PSMA expression. So granted this is a small site, so there may be a bit of partial volume averaging there. But then there was a second focus that was identified on the PSMA PET that was very hard to see on the T2 images. You can see it in retrospect on the diffusion sequences. But this is where the PSMA PET can potentially complement the MRI for initial staging. And additionally, as I said, we can detect heterogeneous PSMA expression. Now the clinical significance of that, particularly initial staging, still remains to be determined. But these are two different patients. So this is just the MIP from a gallium 68 PSMA case where they actually had heterogeneous disease. So you can see very highly PSMA expressing disease in the left prostate gland. And then there is lower expressing disease in the right prostate gland that also extended along the seminal vesicles. In fact, I don't have the axial images, because the main finding was that we found a metastasis of that lower intermediate PSMA expressing disease that was similar in PSMA expression to the areas of extraprostatic extension in the pelvis. So I think there's an interesting potential for looking at PSMA for the phenotype. And so I think a lot of work needs to be done in terms of what that means for the patient, because that really PSMA PET, all the data available has really focused on detection. So can we just see where the prostate cancer sites are? And this is a separate case where this patient had garden variety adenocarcinoma, no concerning features. There have been some reports in the literature for PET suggesting that patients with low PSMA expression have neuroendocrine differentiation, but I don't think that's been fully vetted yet. That's been observed anecdotally. This is a patient with, as I said, normal prostate adenocarcinoma without any aggressive features, and they had low PSMA expression in the left gland, and I would say negative PSMA expression in the left gland, and then some low level expression in the right. And the MRI better characterized the full extent of the disease in the prostate. And this was confirmed on biology. I don't have the, I should have shown the axial slices, but the extent of the prostate cancer matched what you saw on the MRI. Finally, as I said, PSMA PET is getting more and more insensitive. We can detect bone metastases, so this is a DCF PYL scan, and this was, I believe, in the setting of biochemical recurrence. So you can see that there are tiny sclerotic bone lesions that would be very hard to see on a CT without a directed review, but you can clearly see them on the PSMA PET, so very high uptake. So that, I think everyone is very well aware of the fact that that is one of the real advantages of PSMA PET. I will say one of the challenges there is that there's no data in terms of guiding management for such small metastases. So all the clinical trials that the medical oncologists rely on focus on resists, measurable disease, and bone scans for identifying the extent of metastases. There are some studies, so the first study showing the potential for PET directing therapy was using C11 choline to look at, in the setting of biochemical recurrence, to determine whether ADT could be put off by doing oligometastatic directed therapy, and so there are ongoing trials now to look at PSMA PET, whether we can delay the start of ADT by directing therapy to the few metastases that are identified when they are identified, and then they're randomized to systemic ADT if they have more than three metastases. So finally, one comment that I'll make that I didn't have a case for specifically, but given the growing use of PSMA PET, both for initial staging and biochemical recurrence, which are the FDA approved indications for PSMA PET, and honestly, it's not even, it's not stated as biochemical recurrence, so anytime you have a patient with rising PSA, that is the clinical indication for PSMA PET for any of the approved PSMA tracers. So you have some flexibility. You don't have to be limited to the pure definition of biochemical recurrence using clinical trials, but what we are finding is that in cases where it is hard to find the site to biopsy in a patient with a rising PSA, I think MRI, as I said, you guys are gonna hear more about that, it does a fantastic job of directing that biopsy, but when there's a challenge, the PSMA PET could be another troubleshooting or a decision-making tool, and so we are getting requests now to do that in patients who've been, where the clinical suspicion is high, but it's been hard to prove it by biopsy. Okay, so in conclusion, as you know, PSMA PET has very high sensitivity specificity for detecting extended prostate cancer. I do think PSMA PET and MRI are going to be potential complementary modalities in the initial staging, and so I think the selection of patients where that will be the most useful remains to be determined, but clearly PSMA PET is, for the most part, at our institution, is used for the high-risk patients, so that's where you could get more complete staging information to then help with your next treatment decisions and determining the next treatment approach. And then, just as an aside, so we do expect a small number of prostate cancers to have low or negative PSMA expression, so we will, I think as our clinical experience expands, we'll have a better sense of what that means, if anything, but I think this is where using the multi-modality imaging to get complete staging information is going to be really helpful, and the other prostate cancer radiopharmaceuticals, like choline and fluciclovine, actually may still play a role in the setting of PSMA negative disease. So I'd like to acknowledge, we actually have a PSMA PET clinical trial, so I have to acknowledge our funding from the University of Washington Institute for Prostate Cancer Research, and thank you for your attention. Okay, so this is gonna be more of a philosophical discussion on how we balance the risks and the benefits of the MRI pathway. Now, all of you should be aware that the MRI pathway is risk-based. Risk-based means that there's gonna be benefits and that there's going to be harms, and when we think about the benefits, there are four major benefits compared to systematic biopsy, which you should all be familiar with. So the biggest benefit is reduced numbers of biopsies that are undertaken, and reduced diagnosis of indolent cancer, and this relates to the negative predictive value. But there are also benefits related to the positive predictive value, and that's a more accurate diagnosis of the cancer that's there. In other words, it correlates better with the prostatectomy, and there is a non-inferiority of the detection of clinically significant disease. So these are the benefits. So then the question is, what are the harms? So the harms obviously result from the negative results, and that's what we need to focus on. Now, what is the consistency of the harms of prostate MRI? So it turns out that if you look on the left-hand side, you'll see that these are 10 UK centers, and if you look at the consistency of the negative predictive value, what you notice is that the harms are consistent between centers, and you miss about 5% to 10% of cancers. If you look at the right-hand side, those are the false positive harms, and these are 25 mostly US centers. And what you notice is that the biopsy yields for a targeted biopsy is inconsistent and highly variable between centers. So in fact, it turns out that the harms from a false positive is worse than a harms from the false negative, and that's something that we should all be aware of. So what strategy should we use so that we reduce these false positive harms while maintaining the major benefit which comes from the true negatives, right? So what strategies can we use? So we wanna focus on the false positive harm, and the first thing we can do is improve the reading of the prostate MRI. Now, do false positives and false negatives differ? Now, sorry, true positives and false positives differ. Okay, so in fact, this has been looked at by this paper recently, and it's quite obvious. I mean, if a lesion shines out and looks bad, and it looks like a Pirates 405, it's going to be more likely to be a significant cancer. The higher the PSA density, the more likely it is that that is a clinically significant cancer. When you see these indeterminate Pirates 3 cases, they are less likely to be clinically significant cancer. So they do differ in their visual appearance. The other thing you can do is you can adjust the threshold at which you call something positive. Now, traditionally in secondary care, what we do is we say three, four, and five is positive, and therefore, all of these false positives appear, correct, on the right-hand side of the graph, these ones over here, right, all of these appear. Now, what stops us doing this? Nothing, nothing stops us saying four and five is positive. If we did that, we would reduce the total number of false positive harms, but what will be the expense? The expense will be to get more false negatives, and we want to avoid under-diagnosing patients. So how do we pick who we should say three, four, and five, and four and five? It turns out it's all dependent on prevalence, right? So in an outpatient urological population, the prevalence is about 30 to 50%. If you take a screening population, the prevalence is 4%. So in a screening population, don't use three, or four, and five, that's just madness, right? Because there's a 96% chance there's no cancer there, so don't over-call. So it's very much dependent on prevalence. So if you have a screening population, use four and five. If it's secondary care, where the prevalence is much higher, use, sorry, here, use three, four, and five. Okay, so you can see how we're balancing benefits and harms. The other thing you can do is use PSA density. I think I'm the third person to say this today. PSA density is really important. Now, most of you, I take it, are from the US, so you won't be following the European literature, but this is in the latest guidelines of the EAU. Now, I've just deconstructed the table, and on the left-hand side, you can see PI-RAD score and the prevalence of clinically significant disease according to the PI-RAD score, and this is the PSA density. And when you combine the two, you end up with this table. Now, if you look at this table very carefully, what you notice is that if you were going to biopsy three, four, and five, the traditional way of doing it, you avoid about 38% of biopsies, so it's about a third of people that normally avoid a biopsy, and in this particular analysis, 38%. You miss less than 10%. Remember, we said it's going to be less than 10%. Here it is 6%, and the number of false positives is going to be 45%, which is exactly what that US trial also showed. Now, if we decide to change that and look at only people with an elevated PSA density, so threes with an elevated PSA density greater than 0.2, and all the fours and fives, what would happen? If you look very carefully, you can see the number of biopsies avoided increases from 38% to 45%, and the number of false positives also decreases from 45 to 39, so 6% reduction in false positives. Now, what about the missed cancers? That goes up by 3%, but you'll notice it's still below the EAU acceptable limit for missing a cancer. In other words, you're allowed to miss one in 10 cancers, 10%. Why is that figure there? Because trust biopsies miss 30%, right? So we're still a whole lot better than systematic biopsy. Now, what stops us even going even further? In other words, taking a negative case, a PI-RADS 1 and 2 with a high PSA density, and then a PI-RADS 3 with a higher PSA density, including all fours and fives. Look what happens then. You see how the number of biopsies avoided stays exactly the same. You'll notice how the number of missed cancers goes down, so it becomes 6%, 5%, but the number of false positives, in fact, goes down to 39%. So you can see how we can reduce the harm by using these smarter strategies. The other thing we can do is we can optimize the way we do the biopsy. Okay, so traditionally, well, in the USA, not necessarily traditionally, but in the USA, this is what happens. Everybody's gonna get a biopsy. If there is a lesion, you get a systematic plus a targeted biopsy, that's over here. Now that's not what happens in Europe, for example. So in the UK, we tend to use this pathway. If you're negative, you get no biopsy. If you're positive, you get only a targeted biopsy. Only a targeted biopsy, okay? In Europe, it's quite different. In Europe, you see, what they do is this. Negative, no biopsy. Positive, you get a targeted plus a systematic biopsy. Now, this has changed. In 2022, this all changed, and a lot of you may not have noticed. So now we are using this risk-based pathway. So in the risk-based pathway, if you're negative and you have a high risk on the basis of a PSA density, then you get a systematic biopsy. If you are positive, but you're a PI-RADS 3 and you have a low PSA density, you don't get a biopsy anymore. If you're higher PSA density, you do get a biopsy. So you can see this within the latest EAU guidelines. So this is the EAU guidelines here. You can see when an MRI is negative and clinical suspicion is low, e.g. PSA density less than 0.15, omit biopsy based on shared decision-making. So you can see how Europe is moving towards this risk-based stratification of patients. Now, what will be the benefits and the harm in an independent cohort? And this was published recently from Brazil. It's interesting, and I can see Leo in the audience. And so if you look on the right-hand side, you'll see that the risk-based pathway, which is with negative cases getting a biopsy if the PSA density is higher, you can see the number of the benefit, the diagnosis of clinically significant cancer doesn't change, these red guys over here. It's exactly the same. Now, look at the number of patients avoiding it. These blue people are also people with benefit. These are the people who avoided a biopsy. If you look between here and look over here, you can see the risk-based pathway enables you more people to avoid a biopsy. So in fact, the risk-based pathway in an independent cohort in fact reduces harm while maintaining the benefit of diagnosing clinically significant cancer and avoiding a biopsy. So the risk-based pathway is really the way to go. Now, what's wrong with just doing a targeted biopsy? A lot of you will just say, I can see a lesion, I'm just gonna biopsy that. A lot of people just do that. Don't do systematic course. Well, there is a problem. There is a problem because hyper-precise biopsies can lead to grade and stage migration, and in fact may not save lives in the long term. And here's an example. So here's a man with three plus three cancer. An MRI shows you this five millimeter lesion. It's a Gleason four, right? And what you have done is you've upgraded this guy's risk. So suddenly you've turned somebody who was an active surveillance into active treatment, and that is potentially dangerous. Why is that dangerous? Because MRI only sees the worst part of the cancer. The rest of the cancer you don't tend to see, and this is called the Will Rogers effect. So in fact, we need to be careful. We may actually be driving people to surgery and radiation when in fact we should be holding back. So just be very careful about this. So what can we do about this? Well, there's something called focal saturation biopsies. So focal saturation biopsies, you find the lesion and you take a penumbra biopsy through that area. That way you get a much better correlation with the actual tumor grade on a prostatectomy. So we really ought to be thinking about going in that direction. So these are my golden rules about how we start to balance benefits and the risks. So in Europe, not in the USA, but in Europe, if you have a negative MRI and there is no particular elevated risk, either family history or PSA density, a MRI is unjustified, but you would need to use a safety net, right? That wouldn't happen in the USA, but certainly happens all over Europe. You need to risk stratify patients for MRI. In other words, select the right person. Know what the risk is before you call four or five or three, four, and five, because if it's a screening population, it'll be four and five. If it's a secondary screening, urological population, it's gonna be three, four, and five. Account for the patient's tolerance to false results, and that's very important. That's something we didn't talk about. Adjust the threshold to biopsy on PSA density or use some sort of risk calculator, and those are being validated at the moment. And lastly, optimize the way you biopsy and how you evaluate the histological specimen by the route at which you did the biopsy. Thank you very much indeed. Good morning. All right, so I haven't spoken on site for the last three years, so forgive me for my errors. Okay, so prostate MRI has been in use for more than three decades, and the expectations from the MRI has been changing over time, and as outlined by our experts today, the tasks are becoming more difficult. So we were asked to stage in order to diagnose patients in the past. Now we are more into dealing with negative prior biopsies, surveillance, and people are asking us even about prognosis. It is very similar to the evolution of the footballistas, as you can see here. So all of them were world stars, but the things asked from them has been changing continuously. So it is very similar. The difficulty level is going up every year almost. So as mentioned by Dr. Padani and Dr. Otto, prostate MRI has been very commonly used for guiding the biopsies, and this is our experience from NCIV published in the McGill Journal of Medicine. The system was developed by this group, which I'm a proud member of. The key message in this paper was with targeted approach, in comparison to the systematic approach, you detect four times, sometimes higher, clinically significant disease depending on the threshold that you set. In the meantime, in a big group of patients, over 400 patients, when we took them to the prostatectomy, we figure out that the error rate is about 3.5%, as you can see here. So without taking the gland out, we can estimate with minimum error what is going on in the patient's prostate. Well, these results in NCI is very good in NCI, but when you just get out of NCI and when you go into the academic centers, 26 centers from Europe and United States, mainly United States centers, these results are not replicable. If you set different PI-RADS thresholds over here, the cancer detection rates, I will use that term because it is more user-friendly term, cancer detection rate is, in more than half of these patients, cancer detection rate is less than 50%. So this is a big problem, and I can't imagine what's going on outside these academic centers. So prostate MRI has a performance inconsistency, which we need to take care of, and how can we do that? One way to do this is to train people, of course, but if you look at this quality cascade, there are six critical steps, and many people are involved in these steps, as you see. So as radiologists, we are mainly dealing with image quality and these two critical steps, interpretation and data preparation. So in any of these steps, if there is a mistake, then you are one of these centers which are performing less than 50%, very likely. So in the next few slides, I will try to show how can we solve these two steps using artificial intelligence. So we in NCI have seen that problem, even before the New England Journal of Medicine paper, and we decided to build an end-to-end AI model. So you just image your patient, you feed into this model, it's gonna find the prostate, it's gonna segment the zones, it's gonna segment the lesion, and give a high-res score to the lesion. So it's an end-to-end cascaded system. We trained this model in almost like 1,400 patients, and all of these lesions, I contoured them and I gave them high-res scores. I felt very responsible because of this inconsistency. I made this sacrifice, it was painful, but I did it. Anyway, so I'm not gonna talk about the prostate segmentation because everybody is doing that, but more into lesion detection and lesion segmentation. So for this task, we use a 3D unit architecture, and for the high-res categorization, we use a ResNet architecture. So the cancer detection rate of this model was 63%, more than 50% cut line, as you can see. So without any human interface, this AI model was giving us this result. And the false positive rate per patient, which is very important, was 0.44. In every other patient, there is one false positive result. So this is something that you need to look if you are purchasing an AI package. And most of these cancers, most of these lesions were clinically significant cancers anyway, and the false positives were mostly benign pathology abnormalities of prostate, such as high-grade PIN, or atrophy, or inflammation. So the Dice Similarity Coefficient is used as a marker to how good you are segmenting. So just give, I wanna give some information about it. So this is the ground truth, this is the AI model, and the more they overlap with each other, the higher your Dice score is. So if it is closer to one, it is perfect. But in our case, it was 0.36. But it is a good start prime point. So a radiologist can grow that contour, and then it can reduce the workload. We're not so proud of our Dice score, but it is okay. And the PIROS classification accuracy was about 58%. So this model was agreeing with me 58% in five-class categorization. So if you look into the literature, the agreement is low to moderate in PIROS literature. So this is even better than that. So an example from this cohort, so the top row is my contours, the bottom is the AI model. And as you can see, these two clinically significant cancers were perfectly found and delineated by the AI model. Okay, so every two weeks, there is an AI paper in Proced. I am following that very closely, maybe more than many of you guys. But nobody knows where are we now. But our colleagues from UK has done this analysis, and I'm not gonna go into details of that. But you should see some important messages from that paper. Most of the AI papers are not utilizing a test, an unseen test population when they report their results, which is a big flaw. Okay, and most of these, again, papers are not utilizing multi-vendor data. Like, Siemens works in Siemens, fine. GE works in GE, fine. No, it shouldn't be like that, okay? And the other one is none of them, or most of them are not statistically justified. So they are underpowered, overfitted studies, unfortunately. And AI is important with transparency, and most of these models are not publicly available, or if they are available, they are not salvageable. So this is where we are with prostate MRI AI. Okay, well, if you look at my literature, I publish lots of papers in my life, and I couldn't find the paper that I'm so proud of yet, but I'm working on it. But seeing that I wanted to make a change with that last article I have shown you, so I wanted to make this AI model alive. And we call this research translational AI research, which you see very rare in routine life, okay? So for this purpose, we build an AI deploy team. We collaborated with NVIDIA, and this is the team that I'm very proud to be a member of. So in a phone call, I explain them what I want. I want, you know, the MR images to be in pucks. I will push a button, and then I will get my AI results in, like, maybe two minutes. And this is the workflow that one of the scientists has drawn for me. So the DICOM images are going into a parser. Parser is, like, splitting the data into three. Then T2-weighted segmentation of prostate, and all of these three sequences are used for detection and classification. And at the end, you get a DICOM RT structure, and you get some report, okay? So I will show you how it works. But before that, this is the, you know, figure that how it happens. So the radiologist pushes the button. There is, you know, like, a puck server, which is talking to an inference engine, and then the inference engine processes data and pushes back the pucks. So for this, you need a dedicated deployment server, which we build up, and we utilize Clara SDK Deploy for this purpose. And it is totally, like, used for research purposes in our clinic, and it is controlled by certain radiologists. So this is our classical pucks care stream. This is the patient data, and then I choose the sequences, and then I push this button over there, and this is the output of this particular patient. And AI is telling me that there is something here, actually, in the right transitions on distal apex periuretral location. And if you look at the MRI carefully, yes, there is a lesion, but without this, you know, without this, if I give this to half of my, if I give this to my colleagues, I'm sure, like, more than half of them will never detect this lesion. They will say, hey, this is BPH, or this is nothing. But we did a targeted biopsy, and this is the prostate segmentation, smooth edges, very critical for targeted biopsies, as you all know. So the tool told me that this is a PI-RADS 5 lesion. It gave a volume, 1.4 cc, and we did a targeted biopsy, Gleason 3 plus 4 with Kruber form pattern. So again, this is what AI was able to show us. So where are we with that algorithm now? So we are able to use it for point care analysis, and the tool gives you the result in one minute and 20 seconds. So if you are patient enough for that, you are gonna have the virtual Turkbay evaluation of your prostate patient in, like, less than two minutes. And it gives us this RT structure, and we are able to use it for our guided biopsies, radiation planning, focal therapy, printing 3D molds, and for research. So everything, this is open source, this is publicly available. All right, so let's give some more examples. 64-year-old patient, PSA 7.1, and if you look at this prostate MRI, maybe you may call out something here in the transition zone, right? Okay, well, if you look at the T2 of this, it is a PI-RADS 1, right? If you look at the DAA, it's PI-RADS 3, DC is negative, PI-RADS 1, no need for a biopsy, right? But if I send this to the AI model, AI model is showing me something over here. And we did a guided biopsy, and it was four plus three mucinose. We trained this model with PI-RADS 1 lesions. People may criticize us, why you are biopsying PI-RADS 1s, PI-RADS 2s, for this reason. One day we will use it. And this tool is capturing those, okay? All right, another example, 63-year-old African American male, PSA 2.8, aplastic anemia, and dysuria. And if you look at this image, there is a big lesion in the right peripheral zone, typical, we can say, right apical base, PI-RADS 5, EP, right? But I looked at all the images of the patient. FDG PET shows us an uptake, which is suspicious for prostate cancer, right? It shows uptake. CT shows contrast enhancement area here. But I spend extra more time, and I went to the one month prior CT, and there was nothing over there. So within a month, it is very unlikely to develop a tumor like this in your prostate if you know there is something, not something we're wrong in your biology, okay? Well, I revised my report, and I just said, look, unlikely to develop cancer, prostatitis. We gave large-spectrum antibiotics, and four weeks later, that lesion was gone. But if I gave this to the AI, AI is telling me that there is a PI-RADS 5 lesion. But guess what? AI doesn't know the patient's ethnicity. AI doesn't know the patient's PSA. AI doesn't know the patient's aplastic anemia and dysuria functions. And plus, AI doesn't have access to any of these images. So whatever you train the AI with, AI gives you whatever the result, okay? Another example here, radiation therapy patient. Again, you know, am I out of time, or? Three more minutes, four more minutes, okay. Injury time, injury time. Okay, okay, okay, just give me some injury time, okay. Thank you so much. So it's a radiation failure patient, and you know, like T2 negative. There's a area here, and DC shows us the uptake, and yes, if you give this to the AI by parametric MRI-based AI, there is a lesion over there. We never trained this system with DC MRI, but it was able to find. If you take the DC MRI from the equation, most of us will call it negative. This is a research done by one of my fellows here. He's sitting, and you can talk to him for further details. How about the quality? For the quality, we have developed two different AI models. I'm gonna go over them briefly here. If you look at this T2 image, it is looking very good in quality, right? But if you push it into our AI model, which we developed in-house by Mason, who is in the audience today, it gave us non-diagnostic result, and the AI model explains us why. This heat map shows us that over here, there's a gas problem. We trained the AI model with gas, and if you look into the reconstruction of the image, the AI model is telling us further information about why this MRI is non-diagnostic. There is a stair-step artifact in this image, so AI is finding that, and why this should be like this, and what is the harm of this? Well, you can't elision-burden estimation here, and you can't do a guided biopsy, so this MRI is non-diagnostic, and this AI is an explainable AI. And this is the hemorrhage detector, the final thing, and again, this is getting the hemorrhage from the scout images. Katie is in the audience, she developed that, and this is just, you know, like, alerting the technologists. And in, okay, you wanna take a picture? Take a picture. So, in conclusion, okay, so prostate MRI is documented to improve localized cancer management, and performance variation is a problem, and AI can be helpful, and it is promising to assist several tasks, such as segmentation, quality evaluation, and classifications, and this is the team I am so proud of. Thank you so much.
Video Summary
The discussion focuses on the role and advancements of PSMA-PET and PET-MR in initial staging of prostate cancer. PSMA-PET is particularly effective for detecting metastatic disease, outperforming traditional imaging modalities like bone scans and CT. Advances in PET scanner technology and radiopharmaceuticals, such as those providing a higher signal-to-noise ratio, have improved the detection of small prostate cancer sites. Studies have demonstrated that PSMA-PET offers superior specificity and sensitivity, particularly for larger nodal metastases, enhancing staging accuracy. Also discussed is the potential complimentary role of PSMA-PET with MRI, particularly in cases of PSMA-negative or low-expression disease. Furthermore, improvements in MRI reading and biopsy strategies are suggested to reduce false positives and enhance diagnostic accuracy. The introduction of AI in prostate imaging is also highlighted, serving as a tool to improve detection and reduce interpretative variability, thus potentially overcoming issues of inconsistent MRI performance. Lastly, the risk-based MRI pathway is proposed to balance benefits and harms in cancer detection, particularly by using PSA density adjustments to decide biopsy necessity.
Keywords
PSMA-PET
PET-MR
prostate cancer staging
metastatic disease detection
radiopharmaceutical advancements
AI in prostate imaging
MRI diagnostic accuracy
RSNA.org
|
RSNA EdCentral
|
CME Repository
|
CME Gateway
Copyright © 2025 Radiological Society of North America
Terms of Use
|
Privacy Policy
|
Cookie Policy
×
Please select your language
1
English