Performance of a Deep-Learning Neural Network to Detect Intracranial Aneurysms from 3D TOF-MRA Compared to Human Readers.

To study the clinical potential of a deep learning neural network (convolutional neural networks [CNN]) as a supportive tool for detection of intracranial aneurysms from 3D time-of-flight magnetic resonance angiography (TOF-MRA) by comparing the diagnostic performance to that of human readers. In this retrospective study a pipeline for detection of intracranial aneurysms from clinical TOF-MRA was established based on the framework DeepMedic. Datasets of 85 consecutive patients served as ground truth and were used to train and evaluate the model. The ground truth without annotation was presented to two blinded human readers with different levels of experience in diagnostic neuroradiology (reader 1: 2 years, reader 2: 12 years). Diagnostic performance of human readers and the CNN was studied and compared using the χ Ground truth consisted of 115 aneurysms with a mean diameter of 7 mm (range: 2-37 mm). Aneurysms were categorized as small (S; <3 mm; N = 13), medium (M; 3-7 mm; N = 57), and large (L; >7 mm; N = 45) based on the diameter. No statistically significant differences in terms of overall sensitivity (OS) were observed between the CNN and both of the human readers (reader 1 vs. CNN, P = 0.141; reader 2 vs. CNN, P = 0.231). The OS of both human readers was improved by combination of each readers' individual detections with the detections of the CNN (reader 1: 98% vs. 95%, P = 0.280; reader 2: 97% vs. 94%, P = 0.333). A CNN is able to detect intracranial aneurysms from clinical TOF-MRA data with a sensitivity comparable to that of expert radiologists and may have the potential to improve detection rates of incidental findings in a clinical setting.