A deep learning-based approach accurately predicted aneurysm size from magnetic resonance angiography (MRA) images, reported the authors of a Dec. 3 study published in the Journal of Digital Imaging.
Cerebral aneurysm can lead to non-traumatic subarachnoid hemorrhage, which can result in permanent brain damage or death. Aneurysm size and location have been identified as risk factors for this event, but radiologist annotation of these factors is slow and repetitive, wrote Joseph N. Stember, MD, PhD, of Columbia University Medical Center in New York, and colleagues. Geometric computing shortcuts have also proved to be inaccurate.
With this in mind, Stember et al. created a convolutional neural network (CNN) to automatically detect cerebral aneurysms from MRA images and produce size predictions for basilar tip aneurysms.
For their first step, the team trained the CNN on 250 MRA maximum intensity projection (MIP) images, then tested the method on 86 remaining images. In the second step, Stember and colleagues applied the CNN to a separate set of 14 basilar tip aneurysms for size prediction.
The CNN identified aneurysms in 85 of 86 testing set cases, achieving a receiver operating characteristic (ROC) area-under-the-curve of 0.87. In the second step, the team found the CNN basilar tip aneurysm size was, on average, 2.01 millimeters (30 percent) different from the radiologist-traced size. In regard to the CNN aneurysm-predicted area, that measurement differed by an average of 8.1 millimieters2 or 30 percent.
“Our CNN showed high accuracy and AUC values, even with a relatively small pre-augmentation training/validation set of 250 images,” the authors wrote. “The algorithm was in turn able to predict linear size of the aneurysm to within a small enough margin of user predictions to lie within the [Unruptured Intracranial Aneurysm Treatment Score] UICATS interclass size gradations, and to predict the correct trend in aneurysm area in comparison with radiologist segmentation.”
Stember and colleagues wrote an automated aneurysm process can allow for pre-populating of radiology reports for more efficient workflow and may reduce inter-observer variability.
“Ultimately, in order to be more clinically useful, future work will focus on CNNs to predict aneurysm location and measure volume for full three-dimensional (3D) image volumes,” Stember et al. concluded.
Matt joined Chicago’s TriMed team in 2018 covering all areas of health imaging after two years reporting on the hospital field. He holds a bachelor’s in English from UIC, and enjoys a good cup of coffee and an interesting documentary.