Deep U-Net Architecture with Curriculum Learning for Left Atrial Segmentation

Lei Jiang, Yan Li, Yifan Wang, Hengfei Cui, Yong Xia, Yanning Zhang

Abstract: Segmentation of the late-stage gadolinium-enhanced magnetic resonance imaging (LGE-MRI) is a critical step in the ablation therapy for atrial fibrillation (AF). In this work, we propose an end-to-end deep learning-based segmentation method for delineating 3D left atrial (LA) structures in multiple domains. The proposed method uses the 6 layers deep U-Net architecture as the segmentation backbone. Curriculum learning is integrated into the deep U-Net architecture, helping the network learn step by step from easy to difficult scene. We have tested normal and strong version of data augmentation methods, to verify the effect of reducing domain shifts. Other techniques like Fourier-based data augmentation and Swin Transformer Block have also been explored to further improve the segmentation performance. The experimental results demonstrate that the strong version of data augmentation method can reduce the domain shifts and achieve more accurate result, with mean Dice score of 0.881 on the validation set of LAScarQS 2022 challenge. The evaluation results demonstrate our method’s effectiveness on left atrial segmentation in multi-sequence cardiac magnetic resonance (CMR) data.