Abstract Number: Esoc2026a421 an Automatic Classification System for Stroke Mechanisms in Intracranial Atherosclerotic Disease Based on Infarct Topology and Deep Neural Network

Abstract Background and aims Intracranial atherosclerotic disease (ICAD) is a major cause of ischemic stroke. Accurate and rapid identification of stroke mechanisms in symptomatic ICAD patients may guide treatment strategies and improve secondary prevention. We aimed to develop an automatic classification system for stroke mechanisms in symptomatic ICAD using convolutional neural networks (CNNs). Methods Patients with ischemic stroke attributed to 50-99% atherosclerotic stenosis of intracranial internal carotid artery or middle cerebral artery were recruited from 3 teaching hospitals. Presence of artery-to-artery embolism (AAE), hypoperfusion and parent artery occluding penetrating arteries (PAO), as a probable stroke mechanism, were manually labeled based on infarct topology on diffusion weighted imaging (DWI). The infarctions were automatically segmented, and models were constructed based on DWI and infarct mask, employing 3D-ResNet18 and 3D-DenseNet121 architectures with squeeze-and-excitation (SE) blocks. Model performance was evaluated using the area under the receiver operating characteristic curve (ROC AUC). Results Overall, 383 symptomatic ICAD patients recruited from two hospitals were randomly split into training (n=245), validation (n=61), and internal testing (n=77) sets. Among them, 244 (63.7%) patients had AAE, 311 (81.2%) had hypoperfusion, and 166 (43.3%) had PAO. Additionally, 55 patients from another hospital were recruited for external validation. The optimal model, SE-DenseNet121, achieved AUCs of 0.94–0.98 (training), 0.87–0.96 (validation), 0.90–0.97 (internal testing), and 0.89–0.94 (external testing) for independent identification of the three stroke mechanisms. Conclusions The SE-DenseNet121 model demonstrated robust accuracy in identifying stroke mechanisms based on DWI infarct topology in symptomatic ICAD patients, warranting further testing in external cohorts. Conflict of interest Ziqi Li, Yue Gu, Yuan Zhong, Hui Fang, Linfang Lan, Yuhua Fan, Yuming Xu, Thomas W Leung, Qi Dou, Xinyi Leng. Nothing to disclose