Abstract WP61: Automated Large Artery Occlusion Detection IN Stroke Imaging - ALADIN Study

• DOI: 10.1161/str.49.suppl_1.wp61
• Publication Date: 2018-10-08T17:54:39Z
• Authors: Clara M Barreira; Mehdi Bouslama; Diogo C Haussen; Jonathan A Grossberg; Blaise Baxter; Thomas Devlin; Michael Frankel; Raul G Nogueira
• Keywords: 03 medical and health sciences; 0302 clinical medicine; 3. Good health

Introduction: Acute ischemic stroke (AIS) treatment relies on prompt clinical suspicion, neuroimaging findings and stroke neurologists’ referrals, especially for Large Vessel Occlusion (LVO) patients. Recent advances in artificial intelligence have revolutionized the field of computer vision. We propose the use of artificial-intelligence based algorithm for detection of LVOs in AIS setting.

Methods: We performed a multi-center retrospective analysis of CTAs, randomly picked from a prospective cohort of AIS adult patients, with and without LVOs, admitted at comprehensive stroke centers, from 2014 to 2016. An experienced stroke neurologist graded the CTAs for the presence of occlusion and occlusion sites. Concurrently, studies were analyzed by Viz-AI-Algorithm® v3.04 - a Convolutional Neural Network programed to detect MCA-M1 and/or ICA-T occlusions. The primary analysis included ICA-T and/or MCA-M1 LVOs versus more distal occlusions or no LVOs. The secondary analysis included any ICA and/or MCA-M1 and/or M2 LVOs versus more distal occlusions or no LVOs.

Results: Analysis of 500 CTAs is ongoing and will be fully presented at the ISC. Interim results are available in 152 CTAs (Age, 64.1+/-15.7; bNIHSS 16 [IQR, 10-22]; bASPECTS 8 [IQR, 6-10]). Data was enriched for LVO (82.2%). For the primary analysis, the algorithm obtained sensitivity of 0.97 and specificity of 0.52, with a PPV of 0.74 and NPPV of 0.91, and overall accuracy of 0.78. For the secondary analysis (M2 and proximal ICA included), the algorithm obtained sensitivity of 0.92 and specificity of 0.75, with a PPV of 0.92 and NPPV of 0.75, and overall accuracy of 0.88. Maximal running time of the algorithm was under five minutes.

Conclusion: The Viz-AI-Algorithm performs remarkably well for proximal intracranial LVOs. Endeavors on optimization of MCA-M2 LVO detection are being implemented. To the best of our knowledge, this is the first AI-algorithm for detecting intracranial LVOs.