Abstract TMP50: Deep Learning for Hyperdense Middle Cerebral Artery Detection Using High Resolution CT

Abstract

Introduction: While CT angiography (CTA) remains the gold-standard for non-invasive diagnosis of large-vessel occlusion (LVO), accurate identification of potential thrombectomy patients on non-contrast CT (NCCT) has many practical workflow benefits. However, current evidence suggests that the hyperdense MCA sign on NCCT is an insensitive marker of LVO. In this study, we hypothesize that high-resolution NCCT (HR-NCCT) greatly increases the sensitivity of the hyperdense MCA sign and that a deep learning model can successfully identify the majority of LVO patients on HR-NCCT. Methods: A retrospective cohort of LVO and non-LVO patients with HR-NCCT (0.5-1.0 mm slice thickness) and CTA imaging was identified. For each patient, the location of a hyperdense MCA was annotated by an expert board-certified neuroradiologist after cross-referencing both HR-NCCT and CTA data. A convolutional neural network (CNN) was developed for localization of hyperdense MCA from HR-NCCT imaging only. In brief, the CNN is a 3D contracting-expanding network (18 layers; 596,364 trainable parameters) optimized with deep supervision and class weights. All model statistics are reported after aggregating results from a five-fold cross-validation. Results: A total of 97 LVO-positive and 110 LVO-negative cases were identified. Of the LVO cases, 81.4% (79/97) contained a hyperdense MCA upon expert review. Using these 79 positive (visible dense MCA) and 110 negative cases for training, the model demonstrated an overall 90.0% accuracy in hyperdense MCA detection (92.9% sensitivity, 88.2% specificity). Overall sensitivity was similar for M1 (91.7%) and M2-M3 (94.7%) occlusions. Additionally, of the 18 LVO cases labeled without a hyperdense MCA by expert review, 5 cases were identified correctly by the model and visually confirmed to match occlusion location on CTA. Conclusion: A deep learning model can identify 80.4% of all CTA-confirmed LVO patients using only HR-NCCT, with sensitivity approaching human experts. Combined with ASPECTS evaluation, NCCT-based imaging triage alone may be sufficient to identify the majority of candidate thrombectomy patients.