Strategies for Communication and Collaboration in Prostate Cancer Care - Episode 5
Dr Gershman summarizes recent imaging advances and remaining unmet needs for patients with prostate cancer, and envisions how the use of artificial intelligence may impact the field in the future.
Dr. Boris Gershman: The availability and use of MRI has really impacted patient outcomes in the last decade because of the shift in the paradigm of MRI, both from screening to surveillance as well as staging, evaluation, and planning within the local therapy. I think in all those domains, MRI has really improved patient care because it reduces the chance for under- grading, under-staging lesions. It makes the biopsy more accurate. It helps inform treatment planning. For instance, in the screening setting, MRI has reduced the need for biopsy in approximately one-third of men with elevated PSA who have quote unquote normal or negative MRI's. Its's improved the accuracy of biopsy in men who have suspicious lesions and reduced mixed cancer diagnosis, particularly for lesions that may not adequately be sampled by the traditional systematic biopsy technique. And more recent data suggests that this applies, even to transparent needle biopsy not transrectal biopsy, that the utilization of image guidance with MRI enhances the detection of clinically significant cancers and may reduce the detection of clinically insignificant cancers. So in that domain, in the repeat biopsy setting as well as for men on active surveillance, the main utility and benefit of MRI is to reduce under-grading, under-staging, and improve the accuracy of biopsy. And similarly for men contemplating definitive local therapy with either surgery or radiation, MRI has added a level of resolution for local staging but improves treatment planning, counseling for expected outcomes in terms of sexual and urinary function and helps improve patient decision making. So I think without MRI, we're sort of flying blind and MRI has really improved all of those different domains in terms of the prostate cancer screening and treatment spectrum.
As with any technology, there's certainly unmet needs and opportunities for improvement with regards to prostate MRI. I think one of the biggest areas of potential opportunity in the future is to improve the accuracy of prostate MRI for detecting prostate cancer as well as grading it based on imaging. As I mentioned before, there's a discrepancy between the rates of cancer detection according to PI-RADS classification in seminal trials with real-world rates of cancer detection, which tend to be much lower and much more variable. This is likely due to a number of factors, some technical, related to the acquisition of images, some related to interpretation and interim observer variability among radiologists and others related to differences in patient populations. But nonetheless, this represents a big opportunity to improve both the accuracy of MRI for detecting cancer, reducing the need for biopsy in some men and avoiding those negative targeted biopsies for men who have lesions on MRI, but also improving consistency across interpretation of MRIs in the real-world setting. In addition, MRI has potential utility in improving the Gleason grading from just an MRI for tumors. Whether this requires novel imaging techniques or novel interpretation techniques remains to be seen but I see this as the next stage of evolution of MRI to improve its performance and accuracy, both for the detection of cancer as well as for the Gleason grading for imaging of cancer.
The question of how AI may impact my practice and specifically with regards to prostate cancer screening and management, is a really interesting one. There's certainly been a lot of applications of artificial intelligence and machine-learning algorithms everywhere in our day-to-day life, from speech recognition on our phones to car technology to self-driving cars and similarly in healthcare. With regards to MRI, the applications of AI and machine-learning algorithms are really in this category of computer vision applications rather than other AI or machine-learning applications, such as model building, etc. Whether it's evaluation of diagnostic images from pathology slides or prostate MRIs, I think the use cases and barriers are very similar. One potential use case is simple segmentation of suspicious regions. This has been done in prostate cancer pathology slides to aid the pathologists and then find areas of cancer and the use case may be expanded scale and efficiency in terms of supporting radiologists or pathologists in the interpretation of images. The other category of things, which I think is more provocative but a little bit further in the future, is whether these algorithms can, in fact, improve the interpretation of prostate MRI for predicting risk of cancer or other outcomes. There's been some provocative studies, one of which, for instance, examined chest radiographs and used the machine-learning algorithm that was developed to predict life expectancy, something you wouldn't think at firsthand would be possible from a chest radiograph but it looked at indirect science, such as an enlarged cardiac silhouette, breast shadows to suggest gender, underlying lung changes to reflect chronic medical conditions, to predict life expectancy. And so it's quite possible that, in a similar vein, machine-learning algorithms or AI more broadly may be able to find different signals in prostate MRI that may predict meaningful outcomes, such as the probability of cancer being present, extra prosthetic extension, or even clinical outcomes down the road. This is certainly a very provocative and compelling use case but one that I think requires a tremendous amount of work in order to be fully integrated into clinical care and reliable. I think the other category, segmentation assisting radiologists in terms of the interpretation of MRIs, whether radiologists ultimately is the final person making the read is a probably much closer use case and more likely to be integrated to clinical care earlier.