A convolutional neural network (CNN) could assist radiologists with the detection and segmentation of suspicious findings on prostate MRI scans, according to a new study published in Radiology.
“Men suspected of having clinically significant prostate cancer increasingly undergo prostate MRI,” wrote lead author Patrick Schelb, German Cancer Research Center in Heidelberg, and colleagues. “The potential of deep learning to support radiologists in dealing with the increasing imaging volumes requires evaluation.”
Schelb et al. compared the performance of a U-Net CNN to that of radiologists using the American College of Radiology’s Prostate Imaging Reporting & Data System (PI-RADS). The study included data from more than 300 men who underwent MRI scans from May 2015 to September 2016.
The team’s U-Net was trained on data from 250 patients and tested on data from another 62 patients. PI-RADS assessments were performed by a team of eight radiologists. All exams were reviewed prior to the study to ensure high image quality.
Overall, the U-Net performed at a level similar to the radiologists’ PI-RADS assessments. Its sensitivity (88%) and specificity (50%) were comparable to the radiologists (92% and 50%, respectively).
“These findings confirm the hypothesis that this approach can extract salient diagnostic information from prostate MRI,” the authors wrote, adding that their patient cohort was “larger than that in many previously published studies using convolutional neural network in prostate MRI.”
The authors did explain their research had certain limitations. For example, comparing the performance of a CNN to a single radiologist as opposed to a group of radiologists “might provide a better assessment.” Also, though the patient cohort was larger than other previous studies, Schelb observed that an even larger cohort could lead to a better performance from the U-Net.