Deep Learning-Based 3D U-Net Cerebral Aneurysm Detection

Subarachnoid hemorrhage, commonly caused by the rupture of cerebral aneurysms, is a life-threatening condition with high mortality and morbidity. With a death rate of roughly 40%, it is highly desirable to detect aneurysms early and decide about the appropriate rupture prevention strategy. Rotational X-ray angiography is a non-invasive imaging modality and enables diagnostics to detect cerebral aneurysms at an early stage. We propose a variation of the 3D U-Net architecture for the detection and localization of these cerebral aneurysms. This model is enhanced with a knowledge-based postprocessing strategy to minimize the false-positive detections per case. Our suggested method shows similar sensitivity statistics compared to state-of-the-art solutions, with a drastically reduced false-positive rate per patient. The described solution is almost entirely accurate on structures larger than 5 mm in diameter but shows difficulties with smaller aneurysms. We show an F2-score of 0.84 and a false-positive rate of 0.41 on a private test set.