Across several clinical trials, Unfold AI appears to help in prostate cancer treatment selection and follow-up.
Artificial intelligence (AI) may better aid in diagnosing patients with prostate cancer, detecting recurrence, and even guiding treatment with radiotherapy, according to Wayne G. Brisbane, MD.
In a recent conversation between CancerNetwork® and Brisbane, an assistant professor of urology at the University of California, Los Angeles (UCLA) Health, highlighted the use of AI in testing for prostate cancer as well as a few studies that have been conducted on its use in the space.
Results from a retrospective study published in European Urology Open Science, which examined whether AI could be used to predict and map intraprostatic tumor extent, indicated that the model was accurate and efficacious in an independent test set.1 Investigators were able to conclude that the strategy could possibly improve treatment margin definition, thereby decreasing recurrence.
More specifically, after investigating Unfold AI in several clinical trials, a tool that may help in prostate cancer treatment selection and follow-up, identification of tumor extent has increased to 98%, urologists report.2 Additionally, guidance provided by the platform resulted in a change in treatment recommendation 28% of the time, which frequently included changing to a more localized strategy.
Additionally, Brisbane discussed research supporting Unfold AI, and how AI might be used to improve patient care.
Brisbane: AI is being used in prostate cancer in a couple of different ways, mostly on the diagnostic side. In my clinical practice, I use it in men who come in with an elevated prostate specific antigen [PSA], which is our best and most reliable way to screen men for prostate cancer. If you have an elevated PSA, maybe it’s something in the prostate that could just be size or some inflammation, but it could be prostate cancer. [Therefore], we need to figure out exactly what that is for men.
We’re using AI to interpret the MRIs. After a PSA elevation, oftentimes men will get an MRI or a biomarker [assessment], but in my [practice], we use MRI quite a bit. We will use the MRIs, and if it doesn’t look like there’s anything to the naked eye, we’ll have an AI go and take a second look at it. [From there] if there is a lesion on the MRI and we do a biopsy. We're using a machine learning algorithm called Unfold AI that carefully predicts the location and extent of the tumor. We use that to help us perform treatments.
For Unfold AI, we took men who went and got their prostate removed—they elected to have a radical prostatectomy, which cured their cancer. [However], some of the tumors are only encapsulating a small portion of the prostate. Those men’s prostate tumors looked small and [they] were candidates for focal therapy. It’s hard to select those patients because MRI is great, but it’s not perfect. It tends to underestimate tumor volume, and we’re not exactly sure whether it underestimates higher or lower. If we’re only going to treat the tumor, we need to know that for sure.
We took the pathology, and we trained an algorithm on that using MRI and the tract biopsy location. [When] we look at these biopsy cores, often sometimes it feels like you’re playing battleship; you’re going through and dropping these cores at certain locations. We track through the three-dimensional location of all of those cores, and that fed that into the algorithm, as well as men’s PSA demographic features and their Gleason score. We use [those factors] to create these models. It’s accurate. We [assessed it using] thousands of biopsy cores, which we validated against patients here at UCLA and then again at Stanford.
AI is nice in that you can use it anywhere. It can answer many questions, but the limitations are based on the quality of the data that you have for training. I would say that to go forward in the diagnostic space, we are currently trying to create datasets that will help us answer specific questions—for example, can we use imaging and PSA to predict who has prostate cancer? With Unfold AI, we’re looking at both where the tumor is inside the prostate but also whether the tumor spreading outside the prostate. That helps us better select who is a good candidate for certain treatments. As the machine learning algorithms continue to improve, we’re also trying now in medicine to think about our data in a way that helps machine learning algorithms improve. A lot of the current clinical trials are capturing images, pathology, and patient demographics in such a way that they will be able to inform machine learning in the next few years.
I tell patients that there are 3 to 4 phases of any cancer. The first is diagnosis. There’s a lot of potential [for AI] in diagnosis in terms of determining who has cancer. Currently, we’re using PSA, which is a blunt instrument; it doesn’t give you the granular detail that you need. At UCLA, we’re looking at the proteins and RNAs [excreted through urine] to see whether those can be [analyzed] with the power of machine learning or AI algorithm to decipher through all those different urine proteins to detect cancer.
The second thing is risk stratification. Once it is determined that a patient has cancer, we need to know how risky it is. Prostate cancer is a very heterogeneous disease. Some men can live with prostate cancer without any risks of their health. There’s some men who die from prostate cancer; we need to figure out who is in what bucket so we can treat them appropriately. I think that AI can help with that, as well.
The third is going to be treatment. We’re still figuring out exactly how AI can help with treatment. In radiation oncology, they’re using sophisticated technology to aim their beams. We’re doing some clinical trials in focal therapy, where we’re saying, 'Okay if we can predict exactly where the tumor is, can we just ablate that tumor?' AI will open up a new treatment area of trying to just kill the tumor and minimize collateral damage. We have a lot of work to do to figure out if it works. Men who are on those clinical trials are helping us advance the field. I’m very optimistic about it as a [means] of opening up more cancer care with fewer [adverse] effects, which is the goal of all of radiation oncology, urology, and medical oncology.
Finally, there is surveillance. We’ll have to see exactly how AI will help us with surveillance. I think it’ll be important as well.
There’s 3 things: [First], for people who are doing research, always think about how data can be curated to be used by machine learning to solve future questions. Because you have to be able to digitize all these data and think about how patients’ information can be digitized in order to inform machine learning. In my space, that is trying to get imaging together, taking pictures of pathology, and then trying to figure out which patient inputs can be digitized. If it can’t be digitized, then the machine learning algorithms can’t use it.
The second thing is that there are active [options] available currently. For men who are trying to decide between focal therapy vs whole gland therapy, we can predict exactly the extent of their tumor to see if they’re a good candidate or not. Those are applications [for AI] that are currently available. The third thing is that as all doctors continue to look at these different products, we’re starting to notice that that there’s a huge amount of these AI interventions that are coming out and we’re probably going to have to learn with time which ones we can rely on and which need a little bit more work.