Zachary Klaassen: Hi, my name is Zach Klaassen, a urologic oncologist at the Georgia Cancer Center in Augusta, Georgia. I'm delighted to be joined in Europe today by Dr. David Albala, who is the chief of urology at Crouse Hospital in Syracuse, also a member of the Associated Medical Professionals of New York and Syracuse. David, thanks for joining us on UroToday.

David M. Albala: Zach, thanks so much and it's really a pleasure. And I kind of like to talk about a hot topic now in artificial intelligence and molecular imaging, and then we'll talk a little bit about theranostics. I think AI is really here. It's got an impact and it's planting itself in various aspects of all surgical domains, and especially now with PSMA PET imaging used in prostate cancer, I think AI is going to play a really unique role and we're going to try to review that in the next few minutes.

So these are my disclosures, but if we look at where artificial intelligence has been, the first development of computers took place in the 1940s, and it really wasn't until the Dartmouth Workshop in 1950 where we started to see artificial intelligence forming and bringing mathematicians, engineers, economists, and political scientists. We then saw sort of peaks and valleys through the '70s, '80s, and 1990s, things started to remain dormant. But it really wasn't until the late 1990s that we started to see applications of artificial intelligence with IBM, computer Deep Blue beating the chess champion.

NASA rovers were using artificial intelligence in 2004. Watson won Jeopardy in 2011. And then we started to see it with Siri and Alexa. And then now ChatGPT is starting to make its inroads. And in medicine you'll see that across multiple specialties and domains, this is really the hot topic at this point. Really what artificial intelligence focused on initially was improving diagnostic accuracy. We could leverage these machine learning algorithms and disease classifications, and it allowed us to do segmentation, breaking down images, looking at low dose imaging, and then essentially being much more accurate. And we'll show you some examples of that in just a minute.

And it also bred interdisciplinary collaboration. It paved the way for broader adoption, patient care, and both diagnostic and therapeutic outcomes. So if we look at artificial imaging, it's really grown since 2000. But in 2012 was the concept of radiomics that really was introduced and this is the extraction of multiple image features with a high throughput. So essentially you get an image, you segment it, you can extract things from it, and then you can analyze it. And that's really the workflow that was developed.

Now, artificial intelligence and imaging really has four main components. It makes us more efficient. The efficacy is really quite good. It decreases errors. It's reproducible. It allows assistance to the number of radiologists that are doing this. And it increases productivity and mitigates the increased workload that individuals have. So if we look at algorithms, we essentially take data and from this data we can perform intelligent tasks and finding different patterns, recognition, and it allows us then to make a diagnosis. And in some ways, that's what we do as physicians. What we do is we take data, we analyze that data, signs and symptoms, and then we develop guidelines on what a condition should be and how it should be treated.

So essentially what we're doing is we're doing what we do in practice every day. We use data. We use images. We use blood tests. We adopt hypotheses. And then we make a diagnosis and it allows us to understand and recognize data and pattern recognition and helps us make better decisions. So machine learning is a branch of artificial intelligence, and we learn by finding patterns in data and then using those patterns we can predict and make predictions of what a disease patient may have and how successful a potential treatment might be.

So if you look at this little schematic, you can see artificial intelligence and we have two sub-branches of artificial intelligence, machine learning, where you essentially take the data and understand the data. And then deep learning. And deep learning includes neural networks, convolutional networks, and it helps us make more precise diagnosis and understand images. So here's really the difference between machine learning and deep learning. Machine learning uses features that are defined by equations, deep learning, it's a feature extraction and selection and classifications. All of this is done simultaneously. Deep learning is data-driven and it reduces intra-reader variability where machine learning can be quantified using computer programs. And indeed, this just gives you an idea of in medical imaging, essentially taking an image. Here you can see is a CT scan with a lung nodule, passing it through these artificial intelligence systems, whether it's deep learning or machine learning, understanding patterns and textures, and that helps make a diagnosis. You can see with deep learning, there are multiple convoluted neural networks that increase the sensitivity in making the diagnosis much more accurate with deep learning systems. This is some data looking at breast cancer and lung cancer, looking at deep learning systems and the deep learning systems have better performance.

You can see the SDE is the deep learning system up here in the lung and in breast compared to different machine learning systems. You can see that they're just much more accurate in assessing what the disease process is. Radiomics takes image information, it decodes it using a phenotype and allows us to predict what may happen with a particular type of treatment survival time. So it's carrying it one step even further. Not only can we rely on the image to make the diagnosis, but we can predict what the outcome might be and survivability of a particular disease process. So at least in nuclear medicine, we have diagnostic capabilities, we have prognostic capabilities, it enhances image quality, and then it allows us theranostics to allow precise treatment planning. For example, with lutetium-177 PSMA, using this has allowed us to make dose predictions for particular treatments.

So this evolution we can see in nuclear medicine has taken place across the board between machine learning to this deep learning system. And you can see that a lot of this initially started in the 1950s, but here we are in 2020 and 2025, we're using transformers and neural networks to try to predict what these disease processes are and how they're going to respond to particular treatments. So that's the area that's really hot is PET really does lend itself really quite nicely to this imaging functionality, understanding the images, getting much more accurate diagnosis.

And as we all know, this is really becoming the standard of care for patients trying to find metastatic disease. We're trending outside of doing bone scans and CT scans and now using PSMA to try to be much more sensitive. This is an example of looking at liver and spleen segmentation. This is a paper that appeared in the Journal of Radiology in 2020 with deep learning systems and essentially the deep learning systems were much more accurate when you compared this to the standard of truth with a radiologist.

Going across the board, these deep learning systems looking at volume of the liver and the lesions in the liver and the spleen are much more accurate than what a radiologist. Here's another example that appeared in European Radiology, the area under the curve. Here you can see the deep learning system. And here you can see the radiologist. And you can see that the bottom line is this deep learning system showed a 90% sensitivity for classifying liver cancer compared to a 60 to 70% chance with the radiologist. And you can see the true positive and the false positive rates are here. What's amazing is this can be done in 5.6 milliseconds. So it's done efficiently and rather incredibly rapidly across the board. Now in PET imaging, initially these deep learning systems were used for image acquisition and quality assessment, and now we're starting to see tumor delineation, registration and quantification across the board.

You can see that the different radionuclides for PET imaging, many of us are familiar with F18 and gallium. Now we're starting to see even copper being used, but clearly F18 is the most commonly used agent, primarily because of its half-life of 1.8 hours and the positron lead and the detection sensitivity. But here you can see a nodule. Here's a PET image and you can see this nodule here in this prostate cancer patient and it lights up rather dramatically across the board. So we're starting to see this by... We can now merge this anatomical information and then essentially use... There are commercially available software with these PSMA targeting agents that help streamline and assess the metastatic burden in individuals. These are all the papers that have been written looking at distant metastasis and lymph node involvement. The bottom line is there's a lot of work being done with all the different gallium F18 but the sensitivity you can see ranges anywhere from 94 to 97%, and the specificity is 82 to 89% across the board in these different studies.

So you have incredible sensitivity and specificity and the accuracy, the areas under the curve, you can see are really quite high, 0.87, 0.81, and so on and so forth. So I think these new imaging modalities do pose challenges with learning curves. Artificial intelligence can be really a lifesaver in some respects because it's much more accurate. This segmentation, understanding metastatic and intraprostatic cancer is really important and the deep learning systems seem to be doing that really quite well. There are limitations to artificial intelligence implementation and PET scanning with the manufacturing, the analysis, the standardization and training. But I think all of these are tasks that can be overcome. There's a lot of data that's being taken in, being segmented and understood and then coming out with an answer across the board. So obviously there are limitations with the quality of the images, the parameters that you're setting, these different scenarios, and there are ethical and legal issues that need to be discussed.

For example, what happens if the artificial intelligence diagnosis is wrong? That hasn't been worked out yet, and I think that that's still as we go forward... I think the bottom line is artificial intelligence holds real promise in making healthcare human again. It brings the physician closer to the patient. We can lessen the drudgery of data entry that leads to doctor fatigue and steals precious time with our patients. So this is an exciting area. We're just at the beginning of it and I think it's going to be a really terrific ride. So Zach, I think we got a lot to look forward to as we go forward.

Zachary Klaassen: David, just an outstanding summary of the talk that you gave at the 2025 Interdisciplinary GU Cancer Symposium. This was held in June in 2025 in St. Petersburg, Florida. I think absolutely your last statement is exactly where I want to jump off. You started off by saying, "There's so much excitement," and there is. This AI's in every single aspect of medicine this day, whether it be pathology, radiology, et cetera, prognostication. Specific to theranostics and imaging, where do you see the potential really exciting breakthroughs over let's say the next three to five years?

David M. Albala: I think what's going to happen is, we're now at the stage of understanding and interpreting the images in a rapid form, but I think we're going to take it the next step is going to be to prognosticate. With this image, you're going to have this kind of an outcome or this potential survival or you're going to respond to this particular drug in this particular way. So that's where the future lies. I think this is great. It simplifies the job for the radiologists and the urologists that are looking at multiple images all the time. But the real value is what else can we extract out of it? And I think the sky's the limit, quite honestly. Artificial intelligence... I'm the editor for the Journal of Robotic Surgery, and the hottest papers coming in are AI and orthopedics. AI and cardiac surgery. It's in every area and you have to embrace it. And I think that if you don't embrace it, you're going to be standing on the street corner alone.

Zachary Klaassen: Yeah, it's a great point. I think when you look at the theranostics is really where we can make inroads. I think we have multiple agents coming. We know that terbium is coming, lead is coming. Actinium is already here to a degree. Lutetium is just the beginning of it. So it's going to be interesting to see as this data matures, as we feed this data into AI algorithms, just exactly what we said, who's going to do well with X, Y, or Z treatment? It's really going to be an exciting time.

David M. Albala: Well, I'll tell you, the future... I never would've predicted when I was a resident that I'd be talking about artificial intelligence. It just was not even on my roadmap. But that's the beauty of what we do. I never thought I'd be a robotic surgeon. Who would've imagined even a laparoscopic surgeon? I was with Clayman when we took the first kidney out and my residency director said, "I'm not sure that's really a great area to go into," and look at where we are now. Nobody's doing open surgery. I did a polycystic kidney disease... I did bilateral nephrectomies in a patient, and I can't remember the last time I opened a patient. And obviously these kidneys were massive and I thought about maybe trying it robotically, but there is a part of medicine that grounds us a little bit.

Zachary Klaassen: That's right.

David M. Albala: But this is a really exciting area and I think there's so much more to learn and each year we're going to learn much, much more.

Zachary Klaassen: Yeah, absolutely. David, great summary. Thanks for your time and expertise on UroToday. We really appreciate it.

David M. Albala: Thanks so much for having me, Zach.