Prostate cancer researchers unveiled a new AI-powered model that can quickly analyze MRIs to accurately predict how prostate cancer tumors may develop and potentially metastasize over time.
The technology uses a segmentation algorithm to quickly analyze MRIs of prostates and outline—in detail—the contours of any cancerous tumors. The model can then calculate the volume of the tumors it identifies, and based upon those results accurately predict the likely future risk that a cancer will—or will not—metastasize after treatment.
Researchers from the Brigham and Women’s Hospital in Boston published their findings in the journal Radiology in October. In the study, they described how their model examined MRIs of patients with prostate cancer, and accurately identified 85% of the most aggressive prostate tumors.
The model completed its image analysis in a matter of seconds, accurately identifying a tumor’s volume. In the study, the researchers noted a direct correlation between a tumor’s volume and the likelihood a cancer would either recur or metastasize after a patient underwent surgery or radiation—the two most common ways to treat prostate cancer.
Dr. Martin T. King, senior author of the paper, noted that the AI model’s speed and accuracy can offer doctors and patients an important new tool for anticipating how prostate cancers may develop.
“We wanted to know, how can we do more with all the diagnostic information we already have? And can knowing the tumor’s volume help us better anticipate outcomes?” said King. “We were coming at this research from the angle of physicians who treat prostate cancer, and who want to have even more insight about how much cancer a patient actually has, and what that may mean for their prognosis.”
Every year, more than 300,000 men in the U.S. are diagnosed with prostate cancer, which is the second most common cancer among men. Men with localized prostate cancer—tumors that have not metastasized—have a five-year relative survival rate of nearly 100%, according to the National Cancer Institute (NCI). However, men whose prostate cancers have spread or metastasized only have a five-year relative survival rate of about 36%, the NCI notes.
To conduct the study, the researchers used an NVIDIA GeForce RTX 3070 GPU running PyTorch. The team also used the open-source nnUNet algorithm to run image segmentation on the MRI data set.
The researchers trained and tested the model on MRI images of prostate cancer tumors from more than 700 cancer patients. They then compared the model’s ability to identify the shapes of tumors and estimate the volumetric size of tumors with the known five-to-ten-year outcomes of the patients.
They found that it was faster and just as accurate as a human expert in identifying a tumor’s volume and location within a prostate. They also discovered that the model’s estimates of a tumor’s volume were potentially relevant for forecasting—or prognosticating—the likelihood a tumor would remain localized or metastasize. According to the study, larger tumors—as measured by volume—were found to be correlated with a higher likelihood of metastasis or recurrence even after being treated by radiation or surgery.
While the study examined a relatively small number of patient MRIs, the researchers note that the early results are encouraging.
“There are going to be newer algorithms that continue to perform better and better,” King said. “There’s going to be a role for AI because it can do so many things much more quickly [than humans] and it can do them consistently.”
Read the full Radiology article.
Check out additional reporting on the researchers’ work.
Access the nnUNet segmentation algorithm the researchers used in their study on GitHub
