Abstract 268: Validation of an automated method to assess 24‐hour infarct volume in stroke patients

Abstract

Introduction In stroke patients, accurate infarct volume assessment at 24 hours typically requires manual segmentation of lesions that often are not well defined. We aimed to validate an automated machine learning algorithm (MethinksFIV) specifically trained to automatically segment and measure subacute infarcts on non‐contrast CT (NCCT) scans. Methods We retrospectively studied stroke patients with a large vessel occlusion prospectively admitted to a comprehensive stroke center. The final infarct volume was segmented on the 24h NCCT manually (man‐FIV) and with the automated method (auto‐FIV). Both measurements were correlated with clinical outcomes. Results We included 346 patients. On 24h‐NCCT median auto‐FIV was lower than man‐FIV (8.6 ml [2.3‐40.7] vs. 16.9 ml [1.1‐70.8], p<0.001). Auto‐FIV and man‐FIV were highly correlated (r=0.8; p<0.001; mean dice coefficient of 0.61). The Auto‐FIV correlated better than man‐FIV with 24h (r=0.54 vs. r=0.51, both p<0.001) and discharge NIHSS (r=0.59 vs r=0.54, both p<0.001). An increase of 1 ml in auto‐FIV was associated with an increase in the odds of higher ordinal mRS at three months (OR=1.01 95%CI 1.01‐1.02, p<0.001). Auto‐FIV was a better predictor of mRS>2 at three months (OR: 1.02 95%CI 1.01‐1.13, p=0.002) than manual segmentation (OR=1.01, 95%CI 1.00‐1.01, p=0.01). In a multivariable analysis, only auto‐FIV remained a significant predictor of mRS>2 at three months (aOR: 1.03 95% CI 1.01‐1.05, p=0.004). Conclusion An automated method to measure subacute infarct volumes in stroke patients is accurate and correlates well with further clinical outcomes. The method appears to outperform manual segmentation and may be used to facilitate infarct characterization in large clinical trials.