Automatic Myocardium Segmentation in Arterial Spin Labeling Perfusion MRI Using Uncertainty-Aware Mask R-CNN

Anne Oyarzun-Domeño, Verónica Aramendía-Vidaurreta, Gorka Bastarrika, Ana Ezponda, María A. Fernández-Seara, Arantxa Villanueva

Keywords: Mask R-CNN, uncertainty, dropout, Myocardial Blood Perfusion

Abstract: Coronary artery disease (CAD) is a leading cause of cardiovascular morbidity and mortality worldwide. Assessing myocardial perfusion is important to detect potential areas of ischemia in patients with suspected CAD. Arterial spin labeling (ASL) allows non-invasive quantification of myocardial perfusion using arterial blood as endogenous tracer. Segmentation of the left ventricular myocardium is critical in the post-processing for ASL images, but it is challenging due to low signal-to-noise ratio (SNR). This study introduces an automatic myocardium segmentation pipeline including uncertainty awareness, employing Mask R-CNN with dropout layers to capture model uncertainty. Our dataset consists of flow-sensitive alternating inversion recovery (FAIR) ASL images from 16 patients with suspected CAD. Our approach achieves robust segmentation results, with similarity coefficient of 75% and 0.3% misclassification rate. We obtain an 80% correlation with real perfusion values.

Submission Number: 78