Unified multimodal learning for stroke triage: joint detection, scoring, and segmentation of acute ischemic stroke

Timely and accurate triage of acute ischemic stroke patients remains a critical challenge in clinical practice, particularly across heterogeneous imaging protocols and multicenter settings. Existing approaches often focus on isolated tasks such as large-vessel occlusion detection, collateral scoring, or infarct segmentation, leading to fragmented pipelines with limited generalizability and increased latency. In this work, we introduce a unified multimodal framework that jointly integrates non-contrast CT, CT angiography, and CT perfusion under a single encoder-decoder architecture. The system employs symmetry-aware encoders, graph-based vascular modeling, and spatio-temporal fusion, followed by multi-task heads for detection, grading, and segmentation. Trained and validated on multicenter datasets, the framework demonstrates consistent gains over state-of-the-art baselines: AUC improvements of up to +0.05 for LVO detection, a +0.11 increase in quadratic-weighted κ for collateral scoring, and a +5% Dice improvement with reduced volume error for infarct delineation. Beyond technical metrics, the model achieves clinically relevant impact by reducing mis-triage and unnecessary transfers while lowering inference time to under one minute in real-world PACS deployment. Ablation studies further confirm the synergistic benefit of multi-task optimization, which narrows the generalization gap and enhances robustness across unseen centers. Taken together, these results highlight the potential of unified multimodal learning to support fast, reliable, and scalable decision-making in acute stroke care.