Artificial Intelligence in Medicine: Machine Learning Helps with PI-RADS Literature Review

At the Sperling Prostate Center we strive to stay abreast of published studies in our field of prostate MRI. There are so many areas to stay on top of. To name just a few:

• Advances in MRI parameters like Diffusion Weighted Imaging or Dynamic Contrast Enhanced
• Comparing in-bore MRI guidance vs. fusion guidance for prostate biopsies
• Differentiating between prostate cancer (PCa) and other prostate abnormalities
• Ways to shorten scan time without compromising quality, etc.

Another important area is testing and validating the effective use of the Prostate Imaging Reporting and Data System (PI-RADS). PI-RADS is a numerical way to score images for suspicion of PCa. Similar to the BI-RADS scoring system for breast MRI, a value from one to five is assigned to the image. The higher the value, the greater likelihood that clinically significant PCa is present, a form of MRI diagnosis of PCa.

When it was first introduced, there was wide variance in scoring based on reader experience and lack of PI-RADS definitional specifications. Therefore, subsequent versions of PI-RADS were issued with increasing technical specifications plus illustrations as examples in order to boost consistency. The task of studying PI-RADS evolution is global. The work is shared by radiologic centers around the world. Clinicians who conduct multiparametric MRI (mpMRI) prostate scans track and communicate their PI RADS findings through published articles. “Each year, hundreds of scientific studies that report on the diagnostic performance of PI-RADS are published.”^[i] That’s a lot for practicing clinicians to keep up with!

Reviews to the rescue

Fortunately, there are other researchers who perform literature reviews—that is, they define a period of time such as 2018 – 2022, search for all published studies on a narrowly defined topic like MRI diagnosis using PI-RADS, and assign criteria to determine study quality. They then analyze results of each paper that meets their criteria, pool and analyze the data, and submit their review paper for publication. As you might imagine, reading a summary report of say, 33 studies is a great time-saver for those of us who are working with patients on a daily basis. It’s impossible to overstate the value of literature reviews for those of us out in the field, and the debt of gratitude we owe for staying up to date on clinical advances.

We recognize how laborious it is to conduct such a review. This is where Artificial Intelligence (AI) comes in. It can make a huge contribution through the application of Machine Learning (ML). A multicenter interdisciplinary European team recently published a study titled “A Machine Learning Framework Reduces the Manual Workload for Systematic Reviews of the Diagnostic Performance of Prostate Magnetic Resonance Imaging” (Aug. 2023).^[ii] To put ML to the test, they chose to narrowly focus on the diagnostic use of PI-RADS, noting that this PCa scoring system evolves. In turn, this makes it essential for centers like ours to be continually informed as new insights are published.

Putting ML to good use

The authors describe the daunting elements involved in a review/analysis: ”Systematic reviews and meta-analyses are highly resource-intensive, as all the search results from different databases must be checked for eligibility according to the article titles and abstracts. … We hypothesize that machine learning could save valuable workload and facilitate faster updating of the review.” To test their theory, they assigned two readers the job of independently assessing 1585 published candidate papers on the PI-RADS topic based on title and abstract (summary), weeding out those that didn’t make the grade (1103 papers), and ultimately agreeing upon the 482 for which the full text would need to be pulled and read. The team assumed it would take each reader between 30 seconds – 7 minutes to screen each title/abstract (depending on their experience and complexity of content). This could mean anywhere from 13 -150+ hours per reader to just to peruse 1585 candidate articles!

They then designed a ML program to sort through the same number of articles and weed out rejects. Finally, they calculated the potential time savings. In using their ML model, they concluded, “For a review of MRI for prostate cancer diagnosis, this approach reduced the screening workload by about 28%.” This may be conservative, as they note that higher work-saving values have been reported for applications “in other domains.”

As one who has participated in AI research related to prostate cancer, and also as one who finds review articles extremely helpful, I am happy to discover how ML can make life easier for those who conduct the reviews, and indirectly how it can improve our patient services by keeping my staff and I fully on board with improvements and advances made available through professional journals.

NOTE: This content is solely for purposes of information and does not substitute for diagnostic or medical advice. Talk to your doctor if you are experiencing pelvic pain, or have any other health concerns or questions of a personal medical nature.

[i] Nedelcu A, Oerther B, Engel H, Sigle A et al. A Machine Learning Framework Reduces the Manual Workload for Systematic Reviews of the Diagnostic Performance of Prostate Magnetic Resonance Imaging. Eur Urol Open Sci. 2023 Aug 30;56:11-14.
[ii] Ibid.