Comparison of automated and visual DWI ASPECTS in acute ischemic stroke

Abstract Background and purpose To assess intra-and inter-rater agreement of the ASPECTS (Alberta Stroke Program Early CT Score) based on diffusion-weighted MRI and to compare it with fully – automated methods (eASPECTS). Methods DWI-ASPECTS of scans of 96 patients with acute ischemic stroke was rated by 2 experts. Automated methods based on thresholding the affected volumes of a coregistered atlas, and a regression tree learning method were established. Intra-rater, inter-rater and human-rater vs. automated methods agreements were investigated based on the intraclass correlation coefficients (ICC) and Bland Altman plots. Results Intra-rater agreement was good for both raters (ICC of 0.91 and 0.93). Inter-rater agreement was worse (ICC = 0.86) indicating a slight bias between both raters. Agreement with automated methods ranged from 0.81 to 0.87. Root-mean-squared deviation was 0.89 and 0.69 for the human raters and ranged from 0.95 to 1.24 for the automated methods. Conclusions Agreement values are on the same order or higher compared to a literature review of CT-based ASPECTS. Automated methods perform slightly worse than human expert ratings, but they still have enough power to determine the DWI-ASPECTS with good precision in a clinical setting.