Excellent agreement between automated deep learning-based and manual DWI infarct volume measurements in hyperacute stroke

Diffusion-weighted imaging (DWI) lesion volume and infarct growth are important imaging markers in acute ischemic stroke, but manual volume measurement is time-consuming and resource-intensive. Deep learning (DL)-based automated segmentation may facilitate rapid assessment; however, its performance on hyperacute DWI has not been sufficiently assessed. The aim was to evaluate the agreement between DL-based automated and manual infarct volume measurements and to compare their ability to predict clinical outcomes in patients treated with mechanical thrombectomy (MT). Consecutive MT-treated patients (September 2014–December 2019) who underwent DWI at admission and at approximately 24 hours were retrospectively analyzed. Manual infarct volume was measured by stroke neurologists. Automated measurements were obtained using DL-based software. Agreement was assessed using Pearson’s correlation, Bland-Altman analysis, and intraclass correlation coefficients (ICC 2,1). Inter– and intra-rater reliabilities were evaluated in a randomly selected subgroup of 150 patients. Predictive ability for a good outcome at 3 months (modified Rankin Scale score 0–2 or stable/improved from premorbid status) was compared using C-statistics and DeLong’s test. A total of 371 patients (677 DWI scans) were included. Manual and automated measurements showed very strong correlation (r = 0.96) with minimal bias (–1.77 mL). The ICC for manual-automated agreement was 0.959 (95% CI, 0.952–0.965), comparable to inter– and intra-rater ICCs. Agreement remained high across onset-to-imaging times and lesion sizes. Predictive abilities for a good outcome were similar for manual and automated admission DWI volume (C-statistics 0.867 vs. 0.861) and infarct growth (0.859 vs. 0.853). Manual follow-up DWI volume showed slightly better predictive ability than automated measurement (0.880 vs. 0.866). DL-based automated infarct volume measurement shows excellent agreement with experienced clinicians, with predictive performance comparable to manual assessment. Automated DWI-based quantification is reliable and feasible for use in hyperacute stroke management.