Segmentation for Out of Distribution T1$\rho$ Cardiac MRI

Nicholas A. Jacobs, Ganesh Adluru, Jason Mendes, Ravi Ranjan, Andrew Arai, Shireen Elhabian, Edward DiBella

Keywords: MRI, Segmentation, T1, T1ρ

TL;DR: Segmentation models trained on T1 data can segment out-of-distribution T1$\rho$ images of the left ventricle.

Abstract: Prior research has shown that Magnetic Resonance Imaging (MRI) with T1$\rho$ weighted contrast images has the potential to detect disease, such as scar tissue (Han et al., 2014). This makes it a useful imaging modality to help cardiologists diagnose and treat cardiac patients, especially patients with kidney disease who cannot receive contrast. Our work shows that segmentation networks trained on clinical T1 datasets, which are more common and abundant than T1$\rho$ datasets, can be used to segment out-of-distribution T1$\rho$ images from pre-clinical studies given sufficient data augmentation during training.

Submission Number: 118