Beyond Biopsy: A Smarter Way to Grade Prostate Cancer
In This Article
- Traditional Gleason scoring based on prostate biopsy may provide an incomplete view of tumor pathology
- Mass General Brigham investigators are developing a machine learning (ML) model to improve preoperative prediction of surgical pathology
- Investigators are hopeful that the model eventually will play a vital role in clinical decision-making
For decades, the Gleason scoring system has been the standard for assessing prostate cancer aggressiveness. Now Mass General Brigham investigators are developing a machine learning (ML) model that could better support clinical decision-making by providing a more comprehensive picture of tumor burden.
Subscribe to the latest updates from Urology Advances in Motion
Lampros Pantazis, MD, a postdoctoral research fellow at Mass General Brigham Urology, is first author of an abstract describing an ML approach for prostate cancer risk stratification. He and his colleagues found that the model significantly outperformed standard clinical approaches in forecasting prostate cancer Gleason Grade Group (GG) at radical prostatectomy (RP).
“We believe this model has the potential to enhance risk stratification and guidance of prostate cancer management,” Dr. Pantazis says. “We still need to evaluate it in larger cohorts, but this is a first step toward showing its potential utility.”
Dr. Pantazis presented the abstract at the 2026 American Urological Association Annual Meeting.
Addressing limitations of Gleason scoring
Typically, 12 to 20 core samples are taken for a prostate biopsy. Gleason scoring, however, only incorporates the two most prevalent cellular patterns, which are added together to produce the Gleason score (now assigned to one of five Gleason grade groups).
As Adam S. Feldman, MD, chief of Urologic Oncology at Mass General Brigham, notes, biopsies can provide an incomplete view of tumor pathology. As a result, he estimates, 30% to 40% of patients experience a change in GG between biopsy and surgical pathology.
“There’s so much more clinically relevant information in the biopsy that we could use to better characterize the disease,” says Dr. Feldman, senior author of the abstract. “The way we’ve been assessing patients is oversimplified.”
Along with second author Hersh Bendre, MD, a urologic oncology fellow at Massachusetts General Hospital, Drs. Pantazis and Feldman set out to create a model that would improve preoperative prediction of the final GG at surgical pathology. The model was developed utilizing clinical, demographic, and pathology variables that any newly diagnosed prostate cancer patient would have available from the Mass General Brigham transrectal, ultrasound-guided fusion prostate biopsy database.
In an internal retrospective evaluation (the test set comprised 25% of the 535 patients), the investigators found that the model far outpaced both biopsy alone and the National Comprehensive Cancer Network (NCCN) risk groups in predicting prostate cancer GG at RP.
“The biopsy GG and NCCN risk groups are essentially the two main clinical factors for prostate cancer risk stratification,” Dr. Pantazis notes. “So, at least in our cohort, our model seems to give a more accurate estimate of the final Grade Group of the prostate cancer. It also improves the detection of clinically significant prostate cancer (GG2 or higher) compared to biopsy alone, albeit with a little bit of reduced specificity.”
“It ultimately comes down to the physician’s clinical interpretation of the available information,” Dr. Pantazis adds. “This will not be the final decision-making tool. Rather, it will be a way to provide information that isn’t available now to the physician making those treatment decisions.”
Further validations ahead
Before their model can be used to inform clinical decision-making, Drs. Pantazis and Feldman stress that they must:
- Validate the model with larger cohorts from other institutions to ensure generalizability across different populations.
- Conduct prospective validations at Mass General Brigham to compare model predictions with actual patient outcomes, including long-term outcomes.
Drs. Pantazis and Feldman believe that if the model continues to show success, it will play an important role in the clinic one day. For instance, it could reveal that a patient initially slated for active surveillance may in fact need treatment. Or, it could indicate that a patient thought to be intermediate risk may instead be higher risk and need more intensive treatment than anticipated.
“AI algorithms with routine clinical data present opportunities for improved care for our patients and could easily be integrated into clinical care such that clinicians will be able to use them to help assess risk,” Dr. Feldman concludes.
Learn about Mass General Brigham Urology
Learn about the Mass General Brigham Genitourinary Cancer Service