Accurate pericardial fat segmentation from cardiac imaging is critical for diagnosing and managing cardiovascular disorders. In this paper, we describe a unique approach to heart fat segmentation based on the Armed Bandit Generative Adversarial Network (ABGAN). This method combines the capabilities of Generative Adversarial Networks (GANs) with the strategic decision-making of multi-armed bandit algorithms to improve segmentation performance. ABGAN was carefully tested on a broad dataset of cardiac magnetic resonance imaging (MRI) scans and showed considerable gains over existing approaches. Our method obtained 92% accuracy, a Dice similarity coefficient (DSC) of 0.89, and a Hausdorff distance (HD) of 5.2 mm, outperforming classic GAN (86% accuracy), SGAN (0.82 DSC), and DCGAN (7.5 mm HD). In addition, ABGAN quantifies uncertainty, which provides essential insights for clinical decision-making. These findings show ABGAN’s potential for enhanced heart fat segmentation, providing a reliable tool for clinical applications and opening the way for future advances in cardiovascular disease detection and management.