A Benchmarking Study of Deep Learning Approaches for Bi-Atrial Segmentation on Late Gadolinium-Enhanced MRIs

Atrial segmentation from late gadolinium-enhanced magnetic resonance imaging (LGE-MRI) provides essential information for patient stratification and targeted ablation treatment for patients with atrial fibrillation (AF). Automatic segmentation based on deep learning approaches, particularly U-Nets architectures, has been widely used to extract the left atrium (LA) automatically in order to overcome manual segmentation limitations. Unfortunately, a lack of an objective benchmarking study evaluated the efficacy of these approaches including recent development. In this work, we performed a comparative study for bi-atrial segmentation with seven U-net architecture variants. We evaluated the models on segmentation of both LA and right atrium (RA) using the 100 3D LGE-MRI dataset from the University of Utah. Moreover, we explored the domain generalization ability using independently acquired 11 LGE-MRIs from Waikato Hospital, New Zealand. Extensive experiments demonstrate that nnU-Net achieved competitive performance over the other state-of-the-art algorithms, such as ResU-Net, U-Net +  +, Attention U-Net, and Swin UNETR, achieving an average Dice accuracy of 91.5%, an average surface distance of approximately 1 mm, and a 95% Hausdorff Distance of about 4.5 mm for LA and RA cavities on the Utah dataset. The results show that nnU-Net, a plain U-Net architecture, is sufficient for atrial segmentation, and the state-of-the-art algorithms are potentially unnecessary for this particular task, given no significant performance improvements were observed between most of the tested architectures. We hope this work will also provide insights into new model development for medical image segmentation.