LABEL-EFFICIENT CARDIAC VOLUME ESTIMATION IN REAL-TIME MRI USING SPARSE BAYESIAN LEARNING

Abstract

Real-time magnetic resonance imaging (MRI) enables imaging under free-breathing conditions, allowing researchers to study the influence of breathing on the heartbeat. However, this advantage comes at the cost of handling large amounts of data - exceeding 4,500 images per heart - making manual segmentation of ventricular volume highly labor-intensive and costly. To improve label efficiency, it is essential to maximize the information gained from each labeled image. We propose a Bayesian model for predicting ventricular volumes with minimal manual labeling. Unlike standard approaches that treat each slice independently, our model incorporates spatial correlations by enforcing smoothness via a discretized 1D Laplace operator, avoiding strong geometric assumptions. Additionally, we exploit conjugacy to compute the posterior analytically, eliminating the need for sampling or variational inference and allowing for efficient hyperparameter optimization via type-II maximum likelihood. This enables our model to effectively utilize spatial information, improving predictions for slices with few or no labeled images. Evaluation on real patient data demonstrates that our method outperforms the baseline, particularly in data-scarce scenarios, where spatial correlation significantly enhances prediction accuracy. These results highlight the potential of our approach to improve segmentation efficiency in real-time MRI, reducing reliance on extensive manual annotations.