Automatic Detection of Mild Intracranial Stenosis using Time-of-Flight Magnetic Resonance Angiography
Chan Nam Nguyen1, Julia Huck2, Davy C Vanderweyen2
1From the Department of Nuclear Medicine and Radiobiology (C.N.N., J.H., D.C.V.), Diagnostic Radiology (D.C.V., K.W.), Pediatrics (S.C.), Anesthesiology, Faculty of Medicine and Health Sciences, University of Sherbrooke, Sherbrooke, Québec, Canada; Department of Psychology (S.C.), Bishop's University, Sherbrooke, Québec, Canada; Department of Neurology (T.R., J.G.), Miller School of Medicine, University of Miami, Miami, USA; Department of Neurology (M.S.V.E.), Vagelos College of Physicians and Surgeons, Epidemiology (M.S.V.E.), Mailman School of Public Health, Columbia University, New York, USA snd Department of Anesthesiology (P.T.), Faculty of Medicine and Health Sciences, University of Sherbrooke, Sherbrooke, Québec, Canada. chan.nam.nguyen@usherbrooke.ca.
Background And Purpose:
Mild intracranial stenosis (MIS), defined as a 20-50% narrowing of cerebral arteries, represents a potential early biomarker of stroke risk and age-related cognitive impairment. Early detection of MIS may help mitigate downstream neurological consequences. However, manual assessment is time-consuming, while current software-based methods lack the sensitivity to detect subtle luminal narrowing, particularly in tortuous or overlapping arterial segments. This study introduces and evaluates a fully automated method for detecting mild intracranial stenosis using TOF-MRA, using expert manual evaluation as the reference standard.
Materials And Methods:
A total of 312 1.5 T TOF-MRA scans from the Northern Manhattan Study (mean age 71 ± 9 years, 41% men) were used for algorithm development and evaluation. A vascular neurologist graded the presence and severity of intracranial stenosis in five major arteries of the Circle of Willis, the anterior cerebral, internal carotid, middle cerebral, posterior cerebral, and basilar arteries, providing the clinical ground truth. We developed a 3D surface-based automated detection algorithm that localizes stenosis sites and computes NASCET-inspired local and global stenosis percentages using cross-sectional distance and area metrics. The algorithm was applied to the manually graded scans, and its performance was evaluated against expert ratings using sensitivity, precision, specificity, and F1_score.
Results:
The neurologist identified 101 non-stenosis scans (708 arteries), 188 scans with at least one mild stenosis (770 arteries), and 23 scans with at least one moderate-to-severe stenosis (29 arteries). The Hybrid cross-sectional distance metric, which integrates both global and local distance-based stenosis percentages, achieved the highest overall performance in MIS detection, with an average F1_score of 89.98%, sensitivity of 89.10%, precision of 90.87% and specificity of 90.60%. Artery-specific analysis showed that the method was most robust in the anterior cerebral arteries (F1_score = 95.00%), followed by the internal carotid (91.34%) and posterior cerebral arteries (91.71%).
Conclusions:
This work introduces the first fully automated method for detecting mild intracranial stenosis from TOF-MRA, achieving fast, reproducible, and expert-level performance. Our approach not only reduces clinical workload but also enables large-scale population analysis of early-stage stenosis, thereby offering new opportunities to study mechanisms underlying mild intracranial stenosis-related cerebrovascular diseases.
