Abstract WP266: Fully Automated Infarct Detection and Segmentation on Follow-up Non-Contrast CT: External Validation in the AcT Trial

Background: Rapid, accurate infarct identification is vital for stroke management. While follow-up non-contrast CT (NCCT) is routinely used to assess post-treatment infarct volume; its interpretation is hindered by subtle features, reader variability, and time constraints. Manual delineation is labor-intensive and inconsistent, limiting use in large trials and practice. We developed and externally validated a fully automated algorithm for infarct detection and segmentation on follow-up NCCT in the AcT (Alteplase Compared to Tenecteplase) trial, enabling consistent and efficient assessment. Methods: We employed a novel segmentation architecture combining a classic U-shaped convolutional neural network with state-of-the-art denoising diffusion probabilistic models to segment infarcts on NCCT scans. The model was trained on CT scans with manually annotated infarcts in the PRoveIT (Measuring Collaterals With Multi-Phase CT Angiography in Patients With Ischemic Stroke) study and validated using data from the AcT trial, in which patients underwent follow-up non-contrast CT after thrombolysis. Infarct presence and extent were determined by expert readers, blinded to algorithm outputs. Model performance was evaluated using standard diagnostic accuracy metrics and the Dice similarity coefficient. Results: 195 cases in PRoveIT were used for training, 1338 cases (923 infarct-positive) in AcT for testing. In testing data, the fully automated tool achieved sensitivity of 89.6% (95%CI: 87.3-91.9%), specificity of 75.0% (95%CI: 72.9-77.1%), positive predictive value of 88.8% (86.0-91.6%), negative predictive value of 76.7% (95%CI: 73.6-79.8%), and accuracy of 85.1% (82.7-87.5%) for infarct detection. The model achieved fair performance for infarct segmentation for volume calculation (Dice 0.537, 95% CI:0.488-0.585 vs 0.768, 95% CI:0.736-0.790 in training set). Some examples are illustrated in Figure 1 . Conclusions: Our automated infarct segmentation model demonstrated robust performance in detecting infarcts within this clinical trial dataset. Future work will focus on further optimizing the algorithm to improve detection and segmentation of ambiguous and challenging infarct regions.