Purpose: Accurate and prompt diagnosis of proximal large vessel occlusion (LVO) is critical, as endovascular thrombectomy is highly effective in improving patient outcomes in patients with LVO. We assessed the accuracy of an automated LVO-detection algorithm in a cohort of acute ischemic stroke patients.
Materials and Methods: A convolutional neural network (CNN) model developed by Aidoc (Tel Aviv, Israel) was used to detect MCA-M1 and/or ICA occlusions. Retrospective review of stroke cases from the institutional stroke database was performed after approval from Institutional Review Board. A total of 243 patients were included in the study- including 105 known proximal LVOs – Internal Carotid artery (ICA) and Proximal MCA (M1) occlusions- confirmed on conventional angiography. Another cohort of 138 consecutive patients undergoing CTA in the month of August 2019 were included for analysis. The algorithm results for the second cohort were compared with the radiologist’s reads. Sensitivity, specificity and accuracy for occlusion results (Positive versus Negative) as well as Site of occlusion was determined.
Results: The first cohort of 105 known proximal LVOs included 79 M1 and 26 ICA occlusions. The algorithm showed a sensitivity of 92.3% and specificity of 94.9% with an accuracy of 94.3% for identifying the site of occlusion. For detecting occlusion (positive versus Negative) in consecutive patients undergoing CTA, the algorithm showed a sensitivity of 87.6% and specificity of 91.0% with accuracy of 89.3%. Conclusions: The automated algorithm had very high sensitivity and specificity for the detection of LVO, as well as identifying the site of occlusion. This has tremendous potential in the emergency setting as a screening tool to expedite formal diagnosis and improve workflow.