ABSTRACT NUMBER: ESOC2026A2019 AUTOMATED CAROTID WEB DETECTION USING NN-UNET: VALIDATION ON THE MR CLEAN REGISTRY

Abstract Background and aims A carotid web (CaWeb) is a shelf-like fibrous lesion at the carotid bulb that may cause ischemic stroke. CaWeb prevalence in stroke patients is estimated at 0.5-1%, but it is likely higher due to limited awareness and its occurrence in younger individuals. A nn-UNet model (Kuang et al., jnis-2024-021782), a deep learning-based segmentation model, achieved 92% accuracy (AUC=0.92) on 58 CaWeb cases, but lacked validation on an external dataset. This study validates the CaWeb detection model on the MR CLEAN registry (2014-2018). Methods We validated the model on 613 randomly selected CTA scans from acute ischemic stroke patients, including all 32 CaWeb-positive cases. The model segments lumen and possible CaWeb. Diagnostic performance was assessed with Receiver-Operating Characteristic (ROC) analysis. Two volume thresholds were determined: one via Youden's index (maximizing sensitivity-specificity balance) and one enforcing minimum specificity of 75% (while maintaining practical thresholds). Performance was evaluated on cropped scans (3cm around carotid bifurcation), approximating original training conditions. Results ROC analysis (AUC=0.65) yielded thresholds of 6mm3 (Youden's index) and 9mm3 (75% specificity). At 6mm3, accuracy was 72% (sensitivity 72%, specificity 72%, F1 score 21%). At 9mm3, accuracy was 77% (sensitivity 68%, specificity 77%, F1 score 24%). Conclusions The model’s accuracy in this validation study was 77% (AUC=0.65), whereas the accuracy of the original model was 92% (AUC=0.92). The low F1 score reflects the severe class imbalance. Future work will finetune the model on this diverse dataset to improve generalizability and reliability of CaWeb detection in patients with stroke. Conflict of interest All authors: Nothing to disclose Figure 1 - belongs to Conclusions