Â鶹´«Ã½Ó³»­

Â鶹´«Ã½Ó³»­

Explore the latest content from across our publications
Join The Â鶹´«Ã½Ó³»­

Log In

Forgot Password?
Create New Account

Loading... please wait

Abstract Details

Large Vessel Occlusion Identification Using Machine Learning on Computed Tomography Angiography
Cerebrovascular Disease and Interventional Neurology
Cerebrovascular Disease and Interventional Neurology Posters (7:00 AM-5:00 PM)
177
To train a convolutional neural network (CNN) to identify large vessel occlusion (LVO) using computed tomography angiography (CTA).
Patients with acute ischemic stroke due to LVO are at high risk for severe outcomes and benefit from early identification. Machine learning may aid this process, yet studies are limited on automated identification of LVO using CTA.
Stroke-alerted patients from a comprehensive stroke center during the period November 2017–May 2019 were included. Exclusion criteria were missing or poor-quality images, intracranial hemorrhage or implant, or rare pathology including posterior circulation LVO. LVO labels were per chart review or neuroradiologist interpretation. Images were processed with registration, skull removal, intensity adjustment, and generation of 40mm axial maximum intensity projection images (MIP) to optimally depict anterior circulation. Phi-Net (Remedios et al., 2018), a deep CNN implemented using Keras and TensorFlow, was trained with 10-fold cross-validation for binary classification of LVO or no LVO. Dataset overall was balanced, though not necessarily in each fold, and 20% was held out for future test set use.
Among 300 eligible patients, LVOs were 57% left-sided, 21% internal carotid artery, 55% M1, and 24% M2. Training included 240 patients. Mean metrics as 95% confidence intervals for test sets across 10 folds are precision-recall area under the curve (AUC) 0.868±0.095, F1 score 0.856±0.04, and receiver operating characteristic (ROC) AUC 0.919±0.048. At threshold 0.5, accuracy 85%±4%, precision 82.8%±8.1%, recall 90.9%±5.3%, and specificity 81.4%±8.7% were calculated.
We successfully classified LVO presence with high performance. With input of a single preprocessed CTA MIP, our ROC-AUC is 0.92 as compared to previously reported values of 0.89 using multi-phase CTA MIPs (Stib et al., 2020), and 0.88 (Sheth et al., 2019) and 0.86 (Amukotuwa et al., 2019) using 3-dimensional CTAs. With improved performance and full automation, future translation into clinical practice may accelerate stroke triage decisions.
Authors/Disclosures
Sneha Lingam, MD (Atrium Health Wake Forest Baptist)
PRESENTER 		Ms. Lingam has nothing to disclose.
No disclosure on file
No disclosure on file
Stephen W. Clark, MD, PhD (Vanderbilt University Medical School) Dr. Clark has nothing to disclose.
No disclosure on file