E-157 Artificial intelligence detection of cerebral aneurysms using CT angiography – proof of concept. (4th August 2020)

• Record Type: Journal Article
• Authors: Mendes Pereira, V; Cancelliere, N; Begin, G; Donner, Y; Levi, G; Wasserman, E; Lobato Mendes, K; Golan, D; Nicholson, P; Nogueira, R; Krings, T
• Is Part Of: Journal of neurointerventional surgery. Volume 12(2020)Supplement 1
• Page Start: A113
• Page End: A113
• Publication Date: 2020-08-04
• Subjects: Nervous system -- Surgery -- Periodicals; Cerebrovascular disease -- Surgery -- Periodicals; 617.48
• DOI: 10.1136/neurintsurg-2020-SNIS.189
• Languages: English
• ISSNs: 1759-8478

Abstract

Abstract : Introduction: Brain Aneurysms (BAs) are a prevalent vascular disease that may cause a life-threatening intracranial hemorrhage. They can often be missed in CTA and MRAs because the diagnosis requires a very methodological approach. Machine learning algorithms have been used to detect large vessel occlusion and other vascular brain conditions. We developed an algorithm using deep neural network to detect and assist BAs. Methods: We developed an algorithm using 3D convolutional neural network modeled as U-net to detect BAs. We used consecutive positive and negative CTAs in two institutions from 2015–2017. The data was annotated by experienced researchers and checked by an experience neuroradiologist. The algorithm construction used initially 179 CTA datasets containing 230 BAs as a training set. After an initial assessment and algorithm optimization, we use 528 CTAs containing 674 BAs and 2400 normal scans as validation set. We aim to perform a blind test on the algorithm to assess its accuracy on detection of BAs using a test set of 300 positive CTAs with BAs independent of the rupture status and larger than 5 mm and 900 negative scans as controls consecutively selected matched by age and sex. We used ROC curves and Pearson correlation tests to assess the algorithm. Results: We are submitting preliminary results of a blind test of 50 positive CTAs and 150 controls. The algorithm achieved a sensitivity of 92% and a specificity of 94% (AUC 0.983). At the time of the conference, we aim to present the complete analysis and subgroup analysis per location, size and rupture status. Conclusion: The Viz. ai aneurysm algorithm was able to accurately detect the majority of brain aneurysms from our blind test dataset. More importantly, it was also able to report consistently the negative scans. Further training should improve even more accuracy particularly on small aneurysm sizes. Disclosures: V. Mendes Pereira: 2; C; iz.ai, Medtronic, Stryker, Balt, Cerenovous, Phenox. N. Cancelliere: None. G. Begin: 5; C; employee of Viz.ai. Y. Donner: 5; C; employee of Viz.ai. G. Levi: 5; C; Emploee of viz.ai. E. Wasserman: 5; C; Emploee of viz.aiployee of Viz.ai. K. Lobato Mendes: None. D. Golan: 5; C; Emploee of viz.aiploee of viz.aiployee of Viz.ai. P. Nicholson: None. R. Nogueira: 2; C; viz.ai. T. Krings: None.