Prostate cancer is the second most common cancer in men, and almost 300,000 individuals are diagnosed with it each year in the U.S. To develop a consistent method of estimating prostate cancer size, which can help clinicians more accurately make informed treatment decisions, Mass General Brigham researchers trained and validated an AI model based on MRI scans from more than 700 prostate cancer patients.

The model was able to identify and demarcate the edges of 85% of the most radiologically aggressive prostate lesions.

Tumours with a larger volume, as estimated by the AI model, were associated with a higher risk of treatment failure and metastasis, independent of other factors that are normally used to estimate this risk.

Furthermore, for patients who received radiation therapy, the tumour volume performed better than traditional risk stratification for predicting metastasis.

Researchers believe the tool could be used to help clinicians understand a tumour’s aggressiveness, to inform more personalised treatment plans, and to guide radiation therapy.

The study is published in the journal Radiology.

“Al-determined tumour volume has the potential to advance precision medicine for patients with prostate cancer by improving our ability to understand the aggressiveness of a patient's cancer and therefore recommend the most optimal treatment,” said first author David D. Yang, MD, of the Department of Radiation Oncology at Brigham and Women’s Hospital, a founding member of the Mass General Brigham healthcare system.

MRI has improved the ability for clinicians to diagnose prostate cancer and is a routine part of diagnosis and treatment.

While human clinicians can estimate tumour size based on MRI images, these estimates are somewhat subjective and can vary from person-to-person.

To develop a more consistent method of estimating tumour size, the researchers trained an AI model based on MRI images of prostate cancer tumours from 732 patients undergoing treatment at a single centre.

They then investigated whether the AI model’s size estimates were associated with treatment success in the 5-to-10 years following diagnosis.

They showed that the AI model was able to locate and measure around 85% of prostate tumours that had a PI-RADS (Prostate Imaging Reporting and Data System) 5 score within the patient cohort.

The score indicates a very high risk of clinically significant prostate cancer.

The model’s size estimates also showed potential as a prognostic marker: larger tumours were associated with higher risk that prostate cancer would come back, as measured by blood levels of prostate-specific antigen (PSA), or metastasize, both for patients who were treated surgically or with radiation therapy.

“The AI measurement itself can tell us something additional in terms of patient outcomes,” said senior author Martin King, MD, PhD, of the Department of Radiation Oncology at the Brigham.

“For patients, this can really tell them something about what are the chances of cure, and the likelihood that their cancer will reoccur or metastasize in the future.”

In addition to helping clinicians and patients understand their cancer’s aggressiveness, the AI model could also help guide radiation oncologists by pinpointing the tumour’s focal region for more targeted treatment.

It’s also a much faster test compared to currently used methods of predicting prostate cancer aggressiveness, which usually take two weeks or longer to yield results.

AI-informed testing could mean that patients can begin treatment sooner.

Cancer research is a foundational pillar in the care Mass General Brigham provides to patients.

Research, along with the power of the system’s strengths in innovation, education and community engagement, allows Mass General Brigham Cancer to deliver integrated cancer care for all, putting health equity at the centre of that support.

The vision is to provide a comprehensive, integrated and research-informed approach to cancer care, helping patients navigate their entire journey of care, from prevention and early detection to treatment and survivorship.

Looking ahead, the researchers are planning to test their model with a larger, multi-institutional dataset.

“We want to validate our findings, using other institutions and patient cohorts with different disease characteristics, to make sure that this approach is generalizable to all patients,” said Yang.

Source: Mass General Brigham
DOI: 10.1148/radiol.240041