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

  • 2021-04-07 16:23:35
  • Yanwen Li, Luyang Luo, Huangjing Lin, Hao Chen, Pheng-Ann Heng
  • 0

Abstract

The novel coronavirus disease 2019 (COVID-19) characterized by atypicalpneumonia has caused millions of deaths worldwide. Automatically segmentinglesions from chest Computed Tomography (CT) is a promising way to assistdoctors in COVID-19 screening, treatment planning, and follow-up monitoring.However, voxel-wise annotations are extremely expert-demanding and scarce,especially when it comes to novel diseases, while an abundance of unlabeleddata could be available. To tackle the challenge of limited annotations, inthis paper, we propose an uncertainty-guided dual-consistency learning network(UDC-Net) for semi-supervised COVID-19 lesion segmentation from CT images.Specifically, we present a dual-consistency learning scheme that simultaneouslyimposes image transformation equivalence and feature perturbation invariance toeffectively harness the knowledge from unlabeled data. We then quantify boththe epistemic uncertainty and the aleatoric uncertainty and employ themtogether to guide the consistency regularization for more reliable unsupervisedlearning. Extensive experiments showed that our proposed UDC-Net improves thefully supervised method by 6.3% in Dice and outperforms other competitivesemi-supervised approaches by significant margins, demonstrating high potentialin real-world clinical practice.

 

Quick Read (beta)

loading the full paper ...