Support Vector Machine-based Spontaneous Intracranial Hypotension Detection on Brain MRI

Background and Purpose To develop a fully automatic algorithm for the magnetic resonance imaging (MRI) identification of patients with spontaneous intracranial hypotension (SIH). Material and Methods A support vector machine (SVM) was trained with structured reports of 140 patients with clinically suspected SIH. Venous sinuses and basal cisterns were segmented on contrast-enhanced T1-weighted MPRAGE (Magnetization Prepared-Rapid Gradient Echo) sequences using a convolutional neural network (CNN). For the segmented sinuses and cisterns, 56 radiomic features were extracted, which served as input data for the SVM. The algorithm was validated with an independent cohort of 34 patients with proven cerebrospinal fluid (CSF) leaks and 27 patients who had MPRAGE scans for unrelated reasons. Results The venous sinuses and the suprasellar cistern had the best discriminative power to separate SIH and non-SIH patients. On a combined score with 2 points, mean SVM score was 1.41 (+/- 0.60) for the SIH and 0.30 (+/- 0.53) for the non-SIH patients (p < 0.001). Area under the curve (AUC) was 0.91. Conclusion A fully automatic algorithm analyzing a single MRI sequence separates SIH and non-SIH patients with a high diagnostic accuracy. It may help to consider the need of invasive diagnostics and transfer to a SIH center.