Hierarchical Relational Inference for Few-Shot Learning in 3D Left Atrial Segmentation

Abstract

Three-dimensional left atrial (LA) segmentation from late gadolinium-enhanced cardiac magnetic resonance (LGE CMR) images is of great significance in the prevention and treatment of atrial fibrillation. Despite deep learning-based approaches have made significant progress in 3D LA segmentation, they usually require a large number of labeled images for training. Few-shot learning can quickly adapt to novel tasks with only a few data samples. However, the resolution discrepancy of LGE CMR images presents challenges for few-shot learning in 3D LA segmentation. To address this issue, we propose the Hierarchical Relational Inference Network (HRIN), which extracts the interactive features of support and query volumes through a bidirectional hierarchical relationship learning module. HRIN learns the commonality and discrepancy between support and query volumes by modeling the higher-order relations. Notably, we embed the bidirectional interaction information between support and query volumes into the prototypes to adaptively predict the query. Additionally, we leverage prior knowledge of foreground and background information in the support volume to model queries. We validated the performance of our method on a total of 369 scans from two centers. Our proposed HRIN achieves higher segmentation performance compared to other state-of-the-art segmentation methods. With only 5% data samples, the average Dice Similarity Coefficient of the two centers respectively reaches 0.8454 and 0.8110. Compared with other methods under the same conditions, the highest values only reach 0.7012 and 0.6898. Our approach improves the adaptability and generalization of few-shot segmentation from LGE CMR images, enabling precise evaluation of LA remodeling.