Reconstruction of complete cerebral arterial anatomy from non-contrast CT using deep learning for pre-thrombectomy guidance.

BACKGROUND AND PURPOSE: Traditional imaging methods, like CT angiography (CTA), are unable to visualize the occluded vessels in acute ischemic stroke (AIS). We aimed to develop a deep learning based segmentation model for reconstructing the complete cerebral vasculature from non-contrast CT (NCCT) in LVO-AIS patients for endovascular thrombectomy planning. METHODS: A nnU-Net model was trained and validated on retrospectively collected paired NCCT-CTA head images without large vessel occlusions (LVOs) from December 2018 to July 2025 (dataset 1: n = 280, for model training, internal validation and internal test; dataset 2: n = 40, for external validation). Model performance was evaluated using quantitative segmentation metrics, including the Dice Similarity Coefficient (DSC). Additionally, the model was evaluated on NCCT images of LVO-AIS patients from two hospitals, and the segmentation results were verified against post-recanalization DSA by two radiologists (dataset 3: n = 290). RESULTS: The nnU-Net model demonstrated robust segmentation performance, achieving DSC of 0.80 ± 0.04, 0.79 ± 0.04, and 0.79 ± 0.04 on the internal validation, internal test, and external validation sets, respectively. In the clinical evaluation involving LVO-AIS patients, high-quality segmentations were assigned by Rater 1 in 98.9% of cases and by Rater 2 in 98.3% of cases, with substantial inter-rater agreement (Cohen's κ = 0.74; 95% CI, 0.56-0.89). CONCLUSION: This study demonstrated the feasibility of using a deep learning model to segment cerebral vasculature from NCCT images in LVO-AIS patients. The proposed approach may provide complementary anatomical information for thrombectomy planning, but prospective studies are needed to determine its clinical impact.