Dual-Consistency Semi-supervised Learning with Uncertainty Quantification for COVID-19 Lesion Segmentation from CT Images

The novel coronavirus disease 2019 (COVID-19) characterized by atypical pneumonia has caused millions of deaths worldwide. Automatically segmenting lesions from chest Computed Tomography (CT) is a promising way to assist doctors in COVID-19 screening, treatment planning, and follow-up monitoring. However, voxel-wise annotations are extremely expert-demanding scarce, especially when it comes diseases, while an abundance unlabeled data could be available. To tackle the challenge limited annotations, this paper, we propose uncertainty-guided dual-consistency learning network (UDC-Net) for semi-supervised lesion segmentation CT images. Specifically, present scheme that simultaneously imposes image transformation equivalence feature perturbation invariance effectively harness knowledge data. We then quantify uncertainty two forms employ them together guide consistency regularization more reliable unsupervised learning. Extensive experiments showed our proposed UDC-Net improves fully supervised method 6.3% Dice outperforms other competitive approaches significant margins, demonstrating high potential real-world clinical practice. (Code available at https://github.com/poiuohke/UDC-Net ).