External Validation and Systematic Error Characterization of Automated Stroke Lesion Segmentation with Exploratory Spatial-Clinical Modeling

Accurate infarct segmentation is essential for quantitative stroke imaging and reproducible downstream analyses. We performed external validation of a deep learning–based automated stroke lesion segmentation framework on an independent cohort of acute ischemic stroke patients (N=138). Segmentation performance was evaluated using Dice similarity coefficient (DSC), precision, recall, and 95th percentile Hausdorff distance (HD95). The model achieved an overall DSC of 0.78 ± 0.18 and HD95 of 6.9 ± 13.1 mm. Stratified analysis revealed reduced performance in small infarcts (<5 mL; DSC 0.61) and a systematic size-dependent bias in medium-sized lesions. Comprehensive failure case analysis identified boundary uncertainty, partial volume effects, and small-lesion detectability as primary error sources. Exploratory spatial-clinical modeling demonstrated modest associations between lesion volume and admission NIHSS (r=0.198, p=0.030). Incorporating spatial features showed a non-significant trend toward improved discrimination of severe stroke (AUC 0.665 vs. 0.603). These findings confirm robust external performance of automated stroke segmentation while highlighting systematic error patterns and the limited explanatory capacity of imaging-only clinical models. The results support cautious integration of topology-aware imaging biomarkers in stroke outcome research.