A convolutional neural network (CNN) can detect and measure cerebral aneurysms in magnetic resonance angiography (MRA) images, according to findings published in the Journal of Digital Imaging.

“Aneurysm size is a key risk factor for hemorrhage and an important consideration in treatment risk,” wrote Joseph N. Stember, MD, PhD, a radiologist at Columbia University Medical Center in New York City, and colleagues. “Automated aneurysm size determination and flagging of aneurysms above a given threshold size can thus be an important application of CNNs and can allow for pre-populating of radiology reports to improve throughput, in addition to providing a remedy for inter-observer variability in size estimations.”

The authors trained CNN to detect aneurysms from MRA images and found its accuracy was 98.8 percent with an area under the ROC curve of 0.87. The CNN was also used to predict the size of 14 basilar tip aneurysms—known to be especially harmful to patients—and its predictions were different, on average, from radiologist-annotated maximum aneurysmal dimensions by 2.01 mm. This difference, the authors noted, would not have impacted clinical management for many of the patients.

“The present work demonstrates that semantic CNNs can derive aneurysm size,” Stember et al. wrote. “Our CNN is particularly well-suited to this task, since it computes aneurysm presence and location on a per-pixel basis with a multi-resolution approach with a U-net-based architecture.”

The authors also noted their research could “be particularly useful as a first or second adjunct reader for detection of small aneurysms.” This could reduce false-negative interpretations, they noted, or help prioritize urgent unread cases as necessary.