P94 Deep learning-based automated device detection for assessment standardisation in mechanical thrombectomy. (29th August 2022)

• Record Type: Journal Article
• Authors: Guerreiro, H; Nielsen, M; Sentker, T; Schmidt, E; Kniep, H; Fiehler, J; Werner, R
• Is Part Of: Journal of neurointerventional surgery. Volume 14(2022)Supplement 2
• Page Start: A46
• Page End: A46
• Publication Date: 2022-08-29
• Subjects: Nervous system -- Surgery -- Periodicals; Cerebrovascular disease -- Surgery -- Periodicals; 617.48
• DOI: 10.1136/neurintsurg-2022-ESMINT.113
• Languages: English
• ISSNs: 1759-8478

Abstract

Abstract : Introduction: Clinical benefits of mechanical thrombectomy (MT) with stent retriever (SR) has shown irrefutable evidence. Post-operative reconstruction of procedural steps is due to lack of conformity in image documentation often challenging. To simplify complexity of image evaluation, automated image interpretation could be applied. In this way, a relation between procedural stage, devices and thrombolysis in cerebral infarction scoring (TICI) 1 could be assessed. Aim: To prove the feasibility of deep learning algorithm for recognition of SR in digital subtraction angiography (DSA) image data, obtained during MT. We aim to identify SR on a given DSA-Series and whether it corresponds to a Solitaire (Medtronic, Minnesota, USA). Methods: 1784 Series were analysed. These were divided in three classes: no device (n=1614), solitaire (n=80) or other devices (n=170). Deep neural network and multihead-attention block 2 was used. Dimensionality reduction from 2d+t to 2D was achieved by applying a maximum-intensity-projection on each view. Views are encoded using an XCiT based encoder Network 3 . Image representations were fused using a multihead-attention-module and classified through a linear layer. The network was trained in an end-to-end fashion with standard cross entropy as a loss function Results: Regarding presence of SR, 94% of no device series and 89% including device were identified correctly. Solitaire could be identified in most instances (82%), other devices are still classified accurate above chance, 21% were confused with Solitaire . Overall balanced accuracy was 81%. Conclusion: Automated DSA assessment on device presence and differentiation is feasible. Its application on evaluation of wider data-sets should be further evaluated. References: Nielsen M, Waldmann M, Frölich, et al . Deep learning-based automated thrombolysis in cerebral infarction scoring: a timely proof-of-principle study. Stroke 2021 Nov;52 (11):3497–504. Vaswani A, Shazeer N, Parmar N, et al . Attention is all You Need. Advances in Neural Information Processing Systems . 2017;30. Ali A, Touvron H, Caron M, et al . XCIT: Cross-Covariance Image Transformers. Advances in Neural Information Processing Systems . 2021 Dec 6;34. Do you have any conflict of interest to declare? : No