Knowledge of the underlying anatomy of the left atrium can promote improved diagnostic protocols and clinical interventions. Hence, an automatic segmentation of the left atrium on magnetic resonance imaging (MRI) can support diagnosis, treatment and surgery planning of heart. However, due to the small size of left atrium with respect to the whole MRI volume, accurate segmentation of left atrium is challenging. Most of the existing deep learning approaches are based on cropping or cascading networks. In this work, we present a novel deep learning architecture for the segmentation of left atrium from MRI volume which incorporates the residual learning based encoder-decoder network. We introduce a loss function and parameter adjustments to deal with the issue of class imbalance and unavailability of large medical imaging dataset. To facilitate the high quality segmentation, we present a three-dimensional multi-scale residual learning based architecture that maintains coarse and fine level features throughout the network. Experimental results have shown a considerable improvement in segmentation performance by surpassing the current benchmarks (especially the winner of Left Atrial Segmentation Challenge-2018) with fewer parameters compared to the state-of-the-art approaches, thus potentially supporting cardiac diagnosis and surgery without adding any extensive pre-processing of input volumes or any post-processing on the base network’s output.

对左心房的基础解剖学的了解可以促进改善诊断方案和临床干预。因此，在磁共振成像（MRI）上自动分割左心房可以支持心脏的诊断，治疗和手术计划。但是，由于左心房相对于整个MRI体积较小，因此准确分割左心房具有挑战性。现有的大多数深度学习方法大多基于裁剪或级联网络。在这项工作中，我们提出了一种用于从MRI体积分割左心房的新型深度学习架构，该架构结合了基于残差学习的编码器-解码器网络。我们引入损失函数和参数调整来处理类不平衡和大型医学影像数据集不可用的问题。为了促进高质量的细分，我们提出了一种基于三维多尺度残差学习的体系结构，该体系结构在整个网络中保持粗略和精细的功能。实验结果显示，与最新方法相比，参数更少，超过了当前基准（尤其是2018年左心房分割挑战赛的获胜者），分割性能有了显着提高，从而有可能在不进行心脏诊断和手术的情况下提供支持在输入网络上添加任何广泛的预处理或对基础网络的输出进行任何后处理。