DC-net: Dual-Consistency semi-supervised learning for 3D left atrium segmentation from MRI

Abstract

Objective:Left atrial segmentation is very important for the treatment of atrial fibrillation. One factor limiting the automatic segmentation of the left atrium is that training network needs a large amount of labeled data, which is expensive and time-consuming. Using limited labeled data for accurate segmentation is our key concern. Methods: In this work, we propose a novel dual-consistency semi-supervised learning method for left atrium segmentation from 3D MR images. Our framework can effectively leverage limited labeled data and abundant unlabeled data by enforcing consistent predictions under model-level and structure-level perturbations. As for model-level perturbations, we employ a shared encoder and two slightly different decoders. Different decoders can output different predictions. As for structure-level spatial contextual perturbations, two sub-volumes with an overlapping region are randomly cropped, taking as inputs under different spatial contexts. Therefore, the proposed method can maintain the invariance of segmentation results when perturbed by different spatial contexts, and be robust to slight perturbations of networks. Results: Our method are evaluated on the public Atrial Segmentation Challenge dataset. The evaluation metrics of Dice, Jaccard, ASD and 95HD are 90.05%, 82.01%, 1.74 voxel and 7.03 voxel when we use 20% labeled data and 80% unlabeled data. The results show that the proposed method outperforms other exiting semi-supervised methods. Conclusion and Significance: The proposed semi-supervised method can achieve accurate segmentation of left atrium by utilizing limited labeled data and abundant unlabeled data, offering an effective way for doctors to diagnose and treat atrial fibrillation.