E-042 Elevating neuro-interventional radiology education: utilizing machine learning and 3D CT reconstructions to understand vascular anatomy. (23rd July 2022)

• Record Type: Journal Article
• Authors: Isikbay, M; Caton, T; Baker, A; Smith, E; Calabrese, E; Amans, M
• Is Part Of: Journal of neurointerventional surgery. Volume 14(2022)Supplement 1
• Page Start: A97
• Page End: A98
• Publication Date: 2022-07-23
• Subjects: Nervous system -- Surgery -- Periodicals; Cerebrovascular disease -- Surgery -- Periodicals; 617.48
• DOI: 10.1136/neurintsurg-2022-SNIS.153
• Languages: English
• ISSNs: 1759-8478

Abstract

Abstract : Introduction/Purpose: Utilizing appropriate projections during cerebral angiography is critical as it can dramatically alter the visualization of key anatomy. Given the unique nature of these projections, a deeper understanding of neurovascular anatomy is needed. Utilizing post imaging processing of CT angiography can be useful to demonstrate this relationship.Historically a common obstacle for CTA image processing is interference from surrounding osseous anatomy, particularly at the skull base. To address this issue, we present a new deep learning algorithm for the automated removal of osseous structures from CTA studies, resulting in superior visualization of vascular anatomy. This processed data can then be easily used for both pre-procedural planning and educational purposes. Materials and Methods: A 3D deep convolution neural network was trained to perform bone removal using 50 CTAs of the brain. The model was then tested on a randomly selected set of 100 CTA brains. Qualitative segmentation accuracy was performed comparing automated segmentations to a publicly available skull removal algorithm (Horos, horosproject.com) for 3D visualization of the carotid siphons. Quantitative segmentation accuracy was evaluated on a randomly selected subset of 56 CTA brains by comparing automated segmentations to manually corrected segmentations using the Dice coefficient.This data was then used to create MIP and 3D reconstructions to demonstrate the relationship between angiographic projection and key vascular anatomy. Results: After training, the algorithm segmented the carotid siphons more accurately than the Horos bone removal tool in 74 out of 100 CTAs (74%). For the remaining 26 cases the segmentation of the carotids was equivalent between methods.Average Dice overlap between the automated and manually corrected segmentations was 0.97.3D reformats after bone removal were manipulated to clearly demonstrated the relationship between angiographic projections and key vascular anatomy (such as between the Water's projection and the basilar artery, Figure 1 ). Conclusion: The presented algorithm is a rapid and accurate method for identification and removal of the skull from previously acquired CTAs of the head. It outperforms publicly available methods for 3D visualization of skull base vasculature. Using our outlined processing algorithm, pre procedural CTA imaging can help clearly demonstrate the relationship between angiographic projections and key cerebral vascular anatomy. The future implications of such processed data on pre-procedural planning is paramount. Using this methodology, the optimal neuro-angiographic projections can be planned prior to a neuro-interventional procedure, which could, in theory, reduce total procedure time. Disclosures: M. Isikbay: None. T. Caton: None. A. Baker: None. E. Smith: None. E. Calabrese: None. M. Amans: None.