Abstract WP77: Automated Accurate Determinations of Acute Infarct Core Volumes From CT Angiography Using Machine Learning

Abstract

Background: Guideline-based acute ischemic stroke (AIS) evaluations require the use of advanced parenchymal imaging including CT perfusion (CTP) to determine the presence of large vessel occlusion (LVOs) and infarct core. The availability of CTP in all hospitals that receive patients with possible AIS is limited. An accurate method to evaluate possible AIS patients that does not require CTP is urgently needed. Methods: Consecutive patients from 3/2018 - 5/2018 evaluated emergently for possible AIS within 24 hours of onset at our institution were identified. Patients were included if they underwent contemporaneous non-contrast CT, CT angiogram (CTA) and CT Perfusion (CTP) with RAPID (IschemaView) post-processing. A linear Support Vector Machine (SVM) was created using CTA data alone and trained against the CTP-RAPID infarct core volume determinations, which was used as the “gold standard.” Performance parameters were calculated using 10-fold cross-validation. Results: Among 139 subjects, median age was 64 [54-73], 48% were female, 40% were white and 34% were African-American. From this population, 16 subjects (12%) were ultimately diagnosed with AIS. Median time from onset to imaging was 3.9 hours [1.7-10.1], NIHSS 16.5 [8-24], and RAPID infarct core and “at risk” volumes were 35.5 mL [13-72] and 128.5 mL [80-165]. The CTA-based SVM classified AIS versus stroke mimic with excellent discrimination (AUC 0.95 ± 0.08) and accuracy of 89% ± 3%. In patients with AIS, CTA-based SVM estimation of infarct core correlated well with the CTP-RAPID (Spearman's rho 0.82 (p<0.001) and Figure ). Conclusion: Our machine learning algorithm was able to accurately discriminate AIS versus stroke-mimic, and reproduce CTP-based infarct core measurements from CTA imaging alone. A complete AIS neuroimaging evaluation, for LVO as well as infarct core, may be obtainable from CTA, an imaging modality with much broader availability than CTP.