Expanding access to cerebrovascular imaging with routine MRI for pre-surgical assessment of intracranial tumor patients: an AI model development and multi-center validation study

BACKGROUND: Angiography is the gold standard for assessing the relationship between cerebral arteries and intracranial tumors, but its use is limited in low- and middle-income countries for various reasons. Contrast-enhanced T1-weighted MRI (T1C) is much more widely available, yet an accurate vessel segmentation method on this modality has not been established. METHODS: In this multicenter study, 1174 cases from four private institutions and one public dataset were used to train and evaluate an automatic segmentation model for intracranial arteries on T1C. Performance was assessed with Dice similarity coefficient (DSC), recall, and Hausdorff distance (HD) on both internal and external test sets. Prospective clinical validation was performed using paired time-of-flight magnetic resonance angiography (TOF-MRA) to compare our model with the gold standard and to investigate the reasons for discrepancies. RESULTS: In internal test, the model achieved DSC = 0.886 ± 0.015, recall = 0.878 ± 0.033, and HD = 10.220 ± 3.821 mm. External test demonstrated robust generalization across four independent institutions, with DSC ranging from 0.859 to 0.884, recall ranging from 0.885 to 0.802, and HD ranging from 9.867 mm to 13.089 mm. Prospectively, clinical evaluation revealed the proposed model identified all target vessels and covered 88.5% of total vessel visible on TOF-MRA, provided a practical anatomic roadmap, but underestimated vessel diameter by 16.7% due to quality of inputs. Still, experts judged the majority of outputs clinically acceptable for presurgical planning. CONCLUSIONS: The model developed a reliable method to segment vessels on T1C, offering a practical alternative for vascular assessment in low-resource settings.