Aims: Atrial fibrillation is the most common cardiac arrhythmia causing morbidity and mortality. The segmentation of left atrium is very important for image-guided ablation of atrial fibrillation and quantification of left atrial fibrosis. However, manual segmentation is labor-intensive and highly subjective. Therefore, the automatic segmentation of the left atrium is of great significance. This study aimed to design a method based on UNet and Bi-directional Long Short Term Memory (LSTM) network for the automatic segmentation of left atrium from LGE-MRI images.
Methods: Due to the high computational cost and GPU memory consumption of 3D deep learning network, 2D deep learning network was used for learning. The input images of different sizes were normalized using zero-mean normalization and cropped. Then, the data augmentation was implemented by applying rotation, flipping, scaling, etc. The images were input to the network after preprocessing. First, the image features were extracted through the convolution structure of UNet network. Then the resolution of the features was restored by upsampling and interpolation, and cascading was performed. The bidirectional convolution LSTM model was combined to obtain the context information in the z-axis direction, which can effectively avoid serious image gaps. In order to solve the problem of data imbalance caused by the target area with small proportion, a weighted loss function was applied. By 3D reconstruction, the final 3D left atrium segmentation results could be obtained.
Results: The model was trained and validated on the dataset of the MICCAI 2018 challenge. The result showed that the method can effectively retain the interlayer information of the images and avoid serious image gaps on the basis of lower computational cost and memory consumption.
Conclusion: A fully automatic segmentation method based on UNet-BiLSTM has been developed, which can be used for 3D reconstruction of the left atrium from LGE-MRI images.