Deep learning offers similar detection of prostate cancer on MRI compared to prostate imaging reporting and data system (PI-RADS) assessments, according to new research out of Germany.
The AI platform—a custom-trained U-Net algorithm—offered similar sensitivity and specificity for detecting clinically significant prostate cancer compared to PI-RADS and segmented prostate boundaries, as well as radiologists. With more research, the algorithm could help radiologists deal with the increasing number of required MRI interpretations.
PI-RADS was created to help standardize prostate MRI interpretation—a difficult job in any clinical setting, wrote first author Patrick Schelb, with the German Cancer Research Center’s Division of Radiology, and colleagues. But even with the tool, accurate interpretation and interobserver variability is difficult to achieve as the demand for prostate MRI interpretations climbs to “unprecedented levels,” the researchers wrote.
Schelb and colleagues trained their U-Net convolutional neural network (CNN) using data from 250 men who underwent MRIs from May 2015 to September 2016. They tested the CNN on a separate cohort of 62 patients. The team compared the U-Net algorithm to the PI-RADS assessments performed by a team of eight radiologists.
Results showed the CNN achieved an 88% sensitivity and 50% specificity, comparable to the 92% and 50% notched by radiologists, respectively.
The sensitivity did decline, however, when radiologists tried to lower the false-positive rate, wrote Anwar R. Padhani, with the Mount Vernon Cancer Center, U.K., and Baris Turkbey, MD, of the National Institutes of Health Molecular Imaging Program, in a related editorial. The tolerance for false-positive rates depends on patient groups, they noted. In men with potentially clinically significant cancers, that threshold is higher.
Going forward, the pair suggested the U-Net system, along with all other forms of AI, undergo prospective randomized studies performed on external data. They encouraged radiologists to take part in creating the future of their profession.
“We recommend that radiologists engage in the clinical development of AI systems for diagnosis of prostate cancer to meet the increasing demands for MRI-directed diagnosis of prostate cancer,” the editorialists concluded.
Read the entire study in Radiology.