Enhancing multi-domain stability of nnU-Net based LGE-CMR left atrium segmentation using the correct data augmentation strategy

Abstract

Cardiovascular diseases are the leading cause of death globally, and Atrial Fibrillation is a critical risk factor. Late-stage gadolinium-enhanced magnetic resonance imaging (LGE-MRI) is of great value in the ablation therapy for atrial fibrillation. Throughout the years, many fully automatic neural network models have been innovated and applied to the segmentation of LGE-MRI scans. nnU-Net is the state-of-the-art universal framework for medical image segmentation, yet the performance of the network might decline when facing domain shift, which is always the case in clinical conditions, and few work are found that intends to optimize its multi-domain stability in LGE Cardiovascular Magnetic Resonance (LGE-CMR) segmentation. This paper aims to specify an effective data augmentation strategy taking into consider the nature of LGE-MRI and CMR to enhance the multi-domain invariance and robustness of the nnU-Net framework. The strategies are evaluated on a testing dataset from a different domain with the training dataset. The performance metrics include Dice, Average Surface Distance and Hausdorff Distance. The selected augmentation strategy achieved stable and outstanding performance across domains in LGE-CMR segmentation tasks.