A Framework with Transformer-Based Model for Cerebrovascular Stenosis Detection in Magnetic Resonance Angiography.
Authors
Affiliations (8)
Affiliations (8)
- Department of Information Management, Yuan Ze University, Taoyuan, Taiwan.
- Vietnam Aviation Academy, Ho Chi Minh City, Vietnam.
- Innovation Center for Big Data and Digital Convergence, Taoyuan City, Taiwan.
- ZDT Group-YZU Joint Research and Development Center for Big Data, Taoyuan City, Taiwan.
- Department of Medical Imaging, Division of Radiology, Far Eastern Memorial Hospital, New Taipei City, Taiwan.
- Graduate Program in Biomedical Information, Yuan Ze University, Taoyuan City, Taiwan.
- Division of Cardiology, Far Eastern Memorial Hospital, New Taipei City, Taiwan. [email protected].
- Graduate Program in Biomedical Information, Yuan Ze University, Taoyuan City, Taiwan. [email protected].
Abstract
Accurate identification of cerebrovascular stenosis is essential for early stroke prevention and effective clinical management. Magnetic resonance angiography provides non-invasive 3D visualization of cerebral vessels, but reliable automated stenosis detection remains challenging due to anatomical complexity and imaging variability. This study aims to develop an automated, robust, and clinically useful transformer-based deep learning framework for detecting stenosis in 3D brain MRA scans. We propose a framework designed for cerebrovascular stenosis detection. It first automatically localizes the centerlines of all arteries and veins within the 3D MRA volume. Each resulting vessel-centered 3D region is then analyzed and classified as normal or narrowed using our proposed transformer-based model. The model was trained and validated on a manually curated, expert-annotated dataset from Far Eastern Memorial Hospital, Taiwan, to ensure high-quality ground-truth labels. Our proposed framework demonstrated strong and stable performance across five-fold cross-validation. Specially, under the imbalanced data setting, the model achieved an average accuracy of 0.9339, F1-score of 0.7998, AUC of 0.9488, and Precision-Recall AUC of 0.8313-indicating robust discrimination capability and effective detection. The experimental results underscore the capability of the proposed framework as a dependable tool for automated cerebrovascular evaluation. Its superior performance indicates significant utility in clinical settings, supporting early detection and risk reduction for stroke.