Atrial fibrillation (AF) is the most common cardiac arrhythmia. Accurate segmentation of the left atrial (LA) and LA scars can provide valuable information to predict treatment outcomes in AF. In this paper, we proposed to automatically segment LA cavity and quantify LA scars with late gadolinium enhancement Magnetic Resonance Imagings (LGE-MRIs). We adopted nnU-Net as the baseline model and exploited the importance of LA boundary characteristics with the TopK loss as the loss function. Specifically, a focus on LA boundary pixels is achieved during training, which provides a more accurate boundary prediction. On the other hand, a distance map transformation of the predicted LA boundary is regarded as an additional input for the LA scar prediction, which provides marginal constraint on scar locations. We further designed a novel uncertainty-aware module (UAM) to produce better results for predictions with high uncertainty. Experiments on the LAScarQS 2022 dataset demonstrated our model's superior performance on the LA cavity and LA scar segmentation. Specifically, we achieved 88.98\% and 64.08\% Dice coefficient for LA cavity and scar segmentation, respectively. We will make our implementation code public available at https://github.com/level6626/Boundary-focused-nnU-Net.

翻译：心房颤动（AF）是最常见的心律失常。准确分割左心房（LA）及LA疤痕可为房颤治疗预后评估提供关键信息。本文提出基于晚期钆增强磁共振成像（LGE-MRIs）的LA腔体自动分割与疤痕量化方法。我们采用nnU-Net作为基线模型，通过引入TopK损失函数来挖掘LA边界特征的重要性。具体而言，训练过程中将聚焦于LA边界像素，从而实现更精确的边界预测。另一方面，将预测LA边界的距离变换作为额外输入用于LA疤痕预测，为疤痕位置提供边缘约束。我们进一步设计新型不确定性感知模块（UAM），以优化高不确定性预测结果。在LAScarQS 2022数据集上的实验表明，本模型在LA腔体和疤痕分割中均展现优异性能：LA腔体分割Dice系数达88.98%，疤痕分割Dice系数达64.08%。实现代码已开源至https://github.com/level6626/Boundary-focused-nnU-Net。