ABSTRACT NUMBER: ESOC2026A2179 ULTRA-EARLY INTRACEREBRAL HEMORRHAGE DETECTION USING AI-ASSISTED TRANSCRANIAL ULTRASOUND: THE HYPER-AI-SCAN PILOT STUDY

Abstract Background and aims Early differentiation between intracerebral hemorrhage (ICH) and ischemic stroke is critical, as management strategies differ substantially, yet this distinction is unavailable in prehospital and hyper-acute settings. This study assessed the feasibility and diagnostic performance of HYPER-AI-SCAN, an AI-assisted transcranial ultrasound framework for ICH detection. Methods In this single-centre prospective pilot study, 100 patients with acute stroke (33 ICH), confirmed by non-contrast CT, underwent portable transtemporal transcranial B-mode ultrasound using a strictly standardized 5-second sweep protocol applied by three operators with different neurosonology experience; infratentorial hemorrhages were excluded. Baseline characteristics were comparable between groups. Ultrasound videos were preprocessed and sampled using a first-pass policy (20 frames from first 2.5 s). A deep-learning pipeline combined a pre-trained ultrasound foundation model with attention-based multiple-instance learning for patient-level ICH classification. Ultrasound-only and cross-modal (NCCT-guided knowledge distillation) training strategies were evaluated, preserving ultrasound-only inference. Performance was assessed using 5-fold patient-disjoint cross-validation. Results In the ultrasound-only cohort (n = 100), balanced accuracy was 0.74 (95% CI 0.60–0.88), sensitivity 0.59 (0.39–0.79), specificity 0.87 (0.71–1.00), AUROC 0.85 (0.68–1.00), and F1 score 0.64 (0.47–0.81). In patients with paired ultrasound–CT data (n = 89), cross-modal training yielded balanced accuracy 0.80 (0.74–0.86), sensitivity 0.69 (0.58–0.80), specificity 0.91 (0.85–0.97), AUROC 0.89 (0.83–0.95), and F1 score 0.72 (0.63–0.81). Conclusions HYPER-AI-SCAN demonstrates the feasibility of AI-assisted, standardized transcranial ultrasound for reliable ICH detection in a real-world, multi-operator pilot setting, supporting scalable ultra-early and prehospital stroke triage. Conflict of interest nothing to disclose. Figure 1 - belongs to Methods