Excellent agreement between automated deep learning-based and manual diffusion-weighted imaging infarct volume...

Diffusion-weighted imaging (DWI) infarct volume and infarct growth are important imaging biomarkers in patients with acute ischemic stroke and treated with mechanical thrombectomy (MT), but manual measurement is time-consuming and resource-intensive. Deep learning (DL)-based automated segmentation may facilitate rapid and reliable assessment; however, its performance on hyperacute DWI and its prognostic value have not been fully validated.

We retrospectively analyzed consecutive patients with acute ischemic stroke treated with MT between September 2014 and December 2019 who underwent DWI on admission and at approximately 24 h. Infarct volumes were measured manually by stroke neurologists and automatically using a DL-based software. Agreement was assessed using Pearson's correlation, Bland-Altman analysis, and concordance correlation coefficient (CCC). Predictive abilities for good functional outcome at 3 months were compared between manual and automated measurements using C-statistics.

A total of 371 patients (677 DWI scans) were included. Manual and automated infarct volumes showed very strong correlation (r = 0.94 for admission DWIs and 0.97 for follow-up DWIs) with minimal bias (-2.82 mL for admission and - 0.50 mL for follow-up). The CCC for manual-automated agreement was 0.935 for admission and 0.971 for follow-up, comparable to human inter- and intra-rater reliabilities. Agreement remained high across onset-to-imaging times. Predictive abilities for good outcome were similar between manual and automated admission DWI volume (C-statistics 0.867 vs. 0.861) and infarct growth (0.859 vs. 0.853).

DL-based automated infarct volume measurement demonstrates substantial agreement with manual assessment and comparable prognostic performance, supporting its feasibility and reliability in hyperacute stroke management.