Abstract WMP62: Infarct Core Volume Estimation on Non-Contrast Computed Tomography in Late Window Patients Using a Machine Learning Algorithm Software

Abstract

Introduction: A simplified patient selection paradigm can reduce time to reperfusion and widen the eligibility of large vessel occlusions (LVO) for endovascular therapy (EVT). We developed and externally validated a machine learning (ML) algorithm to estimate infarct core volume (ICV) on non-contrast computed tomography (NCCT) in anterior circulation LVOs presenting in the late window (≥ 6-24 hours). Methods: LVOs from prospective databases in US and Spain were included. For development, 2858 stroke activations with NCCT/CT angiography[s1] were included. For validation, consecutive LVOs with admission NCCT/CT perfusion (CTP), post-EVT diffusion-weighted imaging (DWI) within 24 hours, and mTICI ≥2b were included. Neuroimaging experts adjudicated ASPECTS on NCCT and final infarct volume (FIV) on DWI (ground truth). ML algorithm was trained using UNet architecture with ResNet 34 encoder to identify ICV on NCCT. Estimated ICVs (algorithm, CTP [<30% cerebral blood flow], and ASPECTS) were compared to ground truth using correlations (intraclass correlation coefficient [ICC] and Spearman’s [r s ]) and Bland Altman plots. Superiority tests between correlations (to compare modalities) and subgroup analyses were performed. Results: 98 patients (median age, 70 years; IQR:59-80; 51% females) were included for validation. Median time to first imaging was 11.4 hours (IQR:9.0-14.7). Correlations (Table) were moderate with the ML algorithm (ICC:0.59) and poor with ASPECTS (r s :-0.31) and CTP ICV (ICC:0.42;). Bland Altman plots showed that ML algorithm had a mean difference closest to zero (-34.2 mL) while the CTP ICV had a larger mean difference (-45.8 mL). Superiority tests (Table), [s2] ML algorithm correlation showed superiority to both the CTP ( p sup =.018) and ASPECTS ( p sup =.003) correlations. Conclusion: On late window patients, estimation of ICV on NCCT with the ML algorithm seems to be superior to CTP and ASPECTS. Further prospective studies are needed.