Large Vessel Occlusions (LVOs) in the M1 segment are among the most frequent and severe causes of Acute Ischemic Stroke (AIS). In this study, we investigate the use of Artificial Intelligence (AI) to detect these occlusions through Deep Neural Networks (DNN). Although AI-assisted segmentation shows considerable promise in medical imaging, the trade-offs between functional and non-functional performance metrics, particularly in clinical and emergency settings, remain insufficiently explored. To address this gap, we adopt an interdisciplinary approach that integrates medical and engineering perspectives to detect and characterize LVOs in the M1 segment of the middle cerebral artery (MCA). We train a state-of-the-art nnU-Net model on a recent NVIDIA GH200 system using late-phase CT angiography (CTA) images from a retrospective cohort of 198 patients. Our results indicate that segmentation performance is influenced by anatomical factors such as collateral circulation and thrombus location, and predictive quality is not strictly dependent on the specific DNN architecture. The model achieves a sensitivity of 0.93 in identifying the thrombus–vessel interface. Furthermore, analysis of non-functional metrics demonstrates that inference costs can be limited to approximately €0.08 per 1000 patients. These findings support the deployment of GPU-based systems in on-premise hospital environments, providing neuroradiologists with an effective assistive and decision-support tool for the detection of LVOs and the planning of surgical interventions in AIS.
Performance Assessment of AI-Assisted Thrombus and Vessels Segmentation in Acute Ischemic Stroke Patients in Emergency Settings
Minghini, Giada;Miola, Andrea;Calore, Enrico;Schifano, Sebastiano Fabio;Zambelli, Cristian;
2026
Abstract
Large Vessel Occlusions (LVOs) in the M1 segment are among the most frequent and severe causes of Acute Ischemic Stroke (AIS). In this study, we investigate the use of Artificial Intelligence (AI) to detect these occlusions through Deep Neural Networks (DNN). Although AI-assisted segmentation shows considerable promise in medical imaging, the trade-offs between functional and non-functional performance metrics, particularly in clinical and emergency settings, remain insufficiently explored. To address this gap, we adopt an interdisciplinary approach that integrates medical and engineering perspectives to detect and characterize LVOs in the M1 segment of the middle cerebral artery (MCA). We train a state-of-the-art nnU-Net model on a recent NVIDIA GH200 system using late-phase CT angiography (CTA) images from a retrospective cohort of 198 patients. Our results indicate that segmentation performance is influenced by anatomical factors such as collateral circulation and thrombus location, and predictive quality is not strictly dependent on the specific DNN architecture. The model achieves a sensitivity of 0.93 in identifying the thrombus–vessel interface. Furthermore, analysis of non-functional metrics demonstrates that inference costs can be limited to approximately €0.08 per 1000 patients. These findings support the deployment of GPU-based systems in on-premise hospital environments, providing neuroradiologists with an effective assistive and decision-support tool for the detection of LVOs and the planning of surgical interventions in AIS.I documenti in SFERA sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.
