AI Use May Improve Treatment Outcomes in Prostate Cancer

“The machine learning algorithms based on pure clinical data aren’t any better, [but] that’s where prostate cancer algorithms are starting to shine,” said James B. Yu, MD, MHS, FASTRO.

A recent Hot Topics article published in the June issue of Oncology^® explored artificial intelligence (AI) use in prostate cancer and its effects on treatment and patient care outcomes.

CancerNetwork® spoke with James B. Yu, MD, MHS, FASTRO, assistant professor adjunct, Department of Radiation Oncology, Smilow Cancer Hospital at Saint Francis Hospital, and Julian C. Hong, MD, MS, assistant professor, Department of Radiation Oncology, University of California, San Francisco, who were authors of the article. They shared their expertise on advances in AI use to better conduct diagnostic imaging, predict clinical outcomes, evaluate histopathology, and plan treatment.

The researchers highlighted a mirroring of innovations seen in the application of AI to prostate cancer to those happening in medicine. Image analysis and computer advances have been applied to classify prostate pathology and imaging, as well as prediction of outcomes. AI tools may improve practice efficiency in radiation oncology as well as patient-facing tools.

Q: How has the use of AI in the prostate cancer field evolved regarding image classification and analysis? What data might support these advances?

Hong: The question was specifically about image classification and analysis, and that’s been one of the major areas where AI has made the first steps in cancer care—in medicine—and a big part of that is related to the analog developments in AI more broadly. Most of this comes from huge advances in computer vision: training computers to distinguish between a dog and a cat. These same types of approaches are being applied in medical areas—in prostate cancer and, more broadly, in cancer in general. The main image classification areas that come up are things like pathology, radiology, and then on the radiation side, radiation planning. Those are probably the big areas computer vision has been contributing to. A lot of the data that’s being used and trained is coming from academic centers. There have been bigger efforts toward pooling more data, which for all AI applications is a major problem—not just having enough data, but also having biased data. Different fields in medicine are ahead in different things. For instance, for computer vision and bias, dermatology has been very much ahead of the game because one of the natural questions is, “Can you identify a skin cancer from a picture?” As a consequence, there have been studies showing that the data that these algorithms are trained on are often in fair-skinned individuals. They’re a biased dataset, so data quality and fairness are also important.

Yu: The most visible use of AI for prostate cancer is this ArteraAI multimodal AI platform, which was built and validated using largely National Clinical Trials Network data, or NCTN data from the NRG. It highlights the importance of federal funding for research and an unintended but very nice result of all this federal funding. [No one] 30 years ago, when RTOG 9408 was being initiated, would anticipate that we would use these slides for this AI histopathology feature classification system that will be built into a prognostic tool. It’s a neat proof of concept that federal funding of research is important and can lead to a myriad of benefits for society. That’s also an issue for imaging data. The federal government constantly has to decide whether to continue to store all of this data. Are we going to continue to support these data repositories? Hopefully, the answer is yes, and they’ll look at applications like this in the future and say, “We don’t know if it’s [going to] pay off now, but in the future, there’s a very good chance it will.”

Q: How have these algorithms become equipped to predict relevant clinical outcomes in prostate cancer?

Hong: The broader realm of AI is trying to have computers do some form of reproducing informal intelligence with a machine, which is more based on the traditional, original definition. Then machine learning is typically considered like a subset of AI specifically geared toward learning from prior data. As far as trying to predict clinical outcomes in prostate cancer, it’s still a work in progress. That’s across the board, probably one of the more long-standing things.

Decipher Prostate, a genomic risk classifier, was built off a model years ago and has been commercially available for years. Even with how long Decipher has been around, we’re still looking for, robust, high-quality evidence for how we use it. We’re trying to incorporate AI and machine learning into more trials to make these types of predictions because, at the end of the day, we’re trying to deliver better care and improve outcomes for patients. It takes trials to figure those things out, so it’s a little bit of a work in progress. There’s [nothing] out there right now that has that high-level, randomized data, but we’re getting there.

Yu: The way I approach this question is [by taking] the words “machine learning” out and [asking] the question again: “Have our algorithms become equipped to predict clinical outcomes in prostate cancer?” Because if an algorithm that was created without some fancy machine learning tool works, it’s just as good as a machine learning algorithm. We see a lot of papers on machine learning algorithms that are based on clinical data, which is just as good as a multivariate logistic regression. The machine learning algorithms based on pure clinical data aren’t any better, to be quite honest, but it's going back to what Dr Hong was talking about with the image detection/image processing feature recognition. That’s where prostate cancer algorithms are starting to shine. The questions that they’re helping to answer are rudimentary at present: “Do these patients need hormone therapy on top of radiation? What’s the risk of the cancer coming back?” They’re not that much better than existing clinical algorithms. Maybe in the future, they’ll be substantively better.

Hong: Those are 2 key important points there. One is AI tools need to be compared with something simple. AI models, and machine learning, tend to overfit on data, which is that they fit too closely to what they’re trained on, and that can cause a lot of issues that you don’t first see. They should always be compared with something simpler so that we can understand what’s going on behind the scenes. Things are very rudimentary right now. We’re jumping around a bit, but there are opportunities for AI to help us do better. In computational health, in our field, that’s where we should be trying to push things because it’s natural that machine learning or AI can model certain things out, that’s to be expected. It's about where do we go from here? How do we how do we push things to improve outcomes, improve treatments, reduce physician burnout? There are a lot of opportunities that are important next steps.

Q: What AI tools have assisted with assessing prostate cancer histopathology?

Hong: [ArteraAI] is probably one of the more high-profile systems that’s out there, and that uses multimodal data in the sense that it uses some clinical data and combines it with computer vision approaches on histopathology slides. It’s going to be an interesting landscape to watch because it’s one of the clear applications of computer vision up front. A lot of people are working on that problem. Some of the more recent ArteraAI work is application-centric: “Should a patient get androgen deprivation therapy or not?” It’s a rapidly evolving landscape. There’ll be a lot of exciting things. It’s all relative, but probably one of the more mature areas, if you will, as far as AI tools, and at the end of the day, they just need to be validated and evaluated on studies.

Yu: The next area would be using the same feature recognition tools and applying them to other cancers. Sure, it’s much harder than that, but you take essentially the same network and see it extracted from a glioma specimen or anything that needs subclassification would be a ready application. I’m sure it’s being done, and it’s been done.

Hong: This goes back to the data [being] important the clinical problem is important, and the context of those 2 things together is important. That will decide the future of how well these things work and how we can help implement them.