LA-HRNet: High-Resolution Network for Automatic Left Atrial Segmentation in Multi-center LEG MRI

Abstract

Atrial fibrillation has become one of the biggest epidemics and public health challenges, and analysis by the late gadolinium-enhanced magnetic resonance imaging (LEG MRI)is of great clinical importance for its diagnosis and treatment. Deep learning-based methods have achieved great success in left atrial segmentation when the MRI data comes from a specific center. However, since images from multiple centers often show large differences, current left atrial segmentation methods designed for single centers often suffer from significant performance degradation when applied to multi-center images. In this paper, we developed a deep network named LA-HRNet for left atrial segmentation in multi-center LGE MRI based on VoxHRNet, a network used for whole-brain segmentation. We made three improvements over the VoxHRNet to make it suited for left atrial segmentation. First, We propose a feature fusion method capable of generating richer features. Second, we propose feature reuse to fuse the multi-scale features generated in the network with subsequent features. Third, we introduce an auxiliary loss in the network. The experimental results on LAScarQS 2022 dataset show that Our proposed improved model has better performance and realizes stronger generalization ability on the multi-center images.