Ischemic Stroke Segmentation and Net Water Uptake Quantification on Multicenter Non-Contrast CT Using Supervised Target-Domain Adaptation

TL;DR - A multicenter study developed an nnU-Net-based, target-domain-adapted model to segment ischemic stroke lesions and quantify net water uptake directly from non-contrast CT. The approach could reduce reliance on advanced imaging, though segmentation accuracy varied substantially by dataset and lesion size.

• Evaluated retrospectively on 801 patients across four datasets, with fine-tuning on small target-domain subsets from Boston and ISLES.
• For lesions ≥30 mL, median Dice scores were 0.68 on Boston data and 0.56 on ISLES.
• Performance declined when smaller lesions were included: median Dice was 0.54 for acute lesions in Boston and 0.20 against post-treatment infarcts in ISLES.
• Automated net water uptake estimation achieved a mean absolute error of 1.37 percentage points on the Boston cohort.