Abstract WP162: Automatic Detection of Middle Cerebral Artery Plaque on T2-weighted Vessel Wall Imaging

• Published in: Stroke (1970) Vol. 50; no. Suppl_1
• Main Authors: Yu, Yannan, Ravkic, Irma, Xu, Wei-Hai, Li, Ming-Li, Liebeskind, David, Scalzo, Fabien
• Format: Journal Article
• Language: English
• Published: 01-02-2019
• ISSN: 0039-2499; 1524-4628
• DOI: 10.1161/str.50.suppl_1.WP162

Abstract only
Background:
Intracranial vessel wall imaging (VWI) has been reported to be an efficient approach for evaluating ICAD, yet the reliability of intracranial plaque identification may vary depending on the experience of institutions or readers. We aim to automatically detect middle cerebral artery plaque on T2-weighted VWI using machine learning algorithms and thus reduce inter-reader variability.
Methods:
Patients were selected from our institutional vessel wall imaging database if they had acute stroke in the MCA territory on diffusion-weighted imaging (DWI) and MCA plaque seen on VWI, or if they had no history of stroke or transient ischemic attack and no plaque seen on VWI. Bilateral MCAs of all patients were analyzed individually. The cross-sectional images of MCA were graded into a scale of 1-3 based on the confidence of the reader regarding the presence of plaque in the cross-sections: Grade 1, the plaque is not present, used as control group; Grade 2, the plaque may or may not present, not included in further analysis; Grade 3, the plaque is present, used as plaque group. Cross-sections with insufficient image quality result from motion artifacts or with non-cross-sectional images of MCA were excluded. The annotated vessel regions on T2-weighted images were used as input to a classification algorithm based on a nonlinear spectral regression algorithm. The model was evaluated under a crossvalidation in order to predict the presence of plaque in the vessel (Grade 1 versus Grade 3).
Results:
A total of 104 patients with stroke and 27 controls were included, leading to 453 slices with plaque and 246 slices for the control group. After a 3-fold crossvalidation, the model reached an accuracy of 87.2% in detecting correctly the presence of plaque in the vessel regions.
Conclusion:
Our experiments demonstrate that a nonlinear machine learning approach can achieve high performance in detecting presence of plaque in the vessel, which is the first step in further analysis of plaque-related parameters and more objective diagnoses.