Abstract WP259: Automated Quantification of Infarct Growth in Endovascular Thrombectomy Using Deep Learning

Introduction: Infarct growth during endovascular thrombectomy (EVT) predicts functional outcomes in acute ischemic stroke patients. Manual quantification from pre- and post-thrombectomy DWI is time-intensive and subject to inter-rater variability. Current deep learning methods focus on single-timepoint segmentation rather than integrated growth analysis. We developed and validated a dual-channel nnUNet model that simultaneously analyzes paired pre- and post-EVT DWI to quantify infarct growth. Methods: We included 118 paired pre- and post-thrombectomy DWI scans from a prospective cohort study as training cases and 82 scan pairs from two clinical trials as external validation set. Six nnUNet architectures were trained: two single-timepoint models using only baseline (BL) or follow-up (FU) scans, and four dual-channel models (DC 1-4) incorporating paired pre- and post-EVT scans. All models used equally weighted cross entropy- and dice-loss and were trained in a 5-fold cross validation setting. Growth for the BL + FU models was defined as the difference between the pre- and post-EVT segmentations. The dual-channel models targeted segmentation of three possible regions: growth area (expansion on post-EVT), common area (lesion present at both time points), and reduction area (lesion resolved on post-EVT) as demonstrated in Figure 1 . Models differed in region selection and hierarchical training order. We compared the models using median (IQR) Dice Similarity Coefficient (DSC) for growth area segmentation. The best-performing model was then evaluated on an external validation set using volumetric agreement (Bland-Altman analysis) for growth volume between ground truth and predicted segmentations, with the single timepoint BL-model also tested for comparison. Results: Median infarct growth was 29.77 mL (training) and 23.09 mL (validation). DC 4 achieved the highest Dice coefficient (0.750, IQR: 0.647-0.827), significantly outperforming all models except DC 2 ( Figure 2 ). External validation showed excellent volumetric agreement with mean bias of 2.77 mL (95% limits: -40.96 to 46.50 mL), compared to BL-model's bias of -22.81 mL (95% limits: -91.64 to 46.01 mL) (Figure 3). Conclusions: A dual-channel nnUNet model incorporating paired pre- and post-thrombectomy DWI significantly outperforms single time-point approaches for automated infarct growth quantification.