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

Yanwen Li, Luyang Luo+3 more|arXiv (Cornell University)|Apr 7, 2021

COVID-19 diagnosis using AI37 references4 citations

TL;DR

This paper proposes UDC-Net, a dual-consistency semi-supervised learning framework with uncertainty quantification for COVID-19 lesion segmentation in chest CT scans. It enforces image-level transformation equivalence and feature-level perturbation invariance to improve unlabeled data utilization, while combining confidence and consensus uncertainty to guide robust consistency regularization, achieving 77.4% Dice score—6.3% higher than fully supervised V-Net and state-of-the-art performance among semi-supervised methods.

ABSTRACT

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 and scarce, especially when it comes to novel diseases, while an abundance of unlabeled data could be available. To tackle the challenge of limited annotations, in this 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 simultaneously imposes image transformation equivalence and feature perturbation invariance to effectively harness the knowledge from unlabeled data. We then quantify the segmentation uncertainty in two forms and employ them together to guide the consistency regularization for more reliable unsupervised learning. Extensive experiments showed that our proposed UDC-Net improves the fully supervised method by 6.3% in Dice and outperforms other competitive semi-supervised approaches by significant margins, demonstrating high potential in real-world clinical practice.

Motivation & Objective

• To address the challenge of scarce, expert-annotated voxel-wise labels for COVID-19 lesion segmentation in CT scans.
• To improve semi-supervised learning efficiency by leveraging abundant unlabeled CT data.
• To enhance model robustness by quantifying and integrating dual uncertainty (confidence and consensus) into consistency regularization.
• To develop a method that generalizes well to real-world clinical settings with irregular and ambiguous lesion contours.
• To outperform existing semi-supervised segmentation methods on a large-scale COVID-19 CT dataset.

Proposed method

• Proposes a dual-consistency learning scheme that enforces image-level transformation equivalence—ensuring predictions remain consistent under input transformations—using a multi-augmented view of the same scan.
• Implements feature-level perturbation invariance by feeding perturbed features from the encoder to auxiliary decoders, which must produce consistent predictions with the main decoder.
• Introduces a multi-decoder V-Net architecture where predictions from multiple decoders are used to compute two forms of uncertainty: confidence uncertainty (mean of predictions) and consensus uncertainty (standard deviation of predictions).
• Uses a learned uncertainty indicator function to filter out low-confidence and high-variability (inconsistent) predictions during consistency regularization, improving training stability.
• Applies a supervised loss on labeled data and a consistency loss on unlabeled data, with uncertainty weighting to suppress noisy or uncertain samples.
• Employs a 3D multi-decoder U-Net backbone with consistency regularization in both input and feature spaces to learn robust, invariant representations.

Experimental results

Research questions

• RQ1 Can dual-consistency learning—simultaneously enforcing input and feature space invariance—improve semi-supervised lesion segmentation in low-resource settings?
• RQ2 Does uncertainty quantification based on multi-decoder outputs enhance the reliability of consistency regularization in semi-supervised segmentation?
• RQ3 Can the integration of confidence and consensus uncertainty lead to better generalization and reduced overfitting on ambiguous lesion boundaries?
• RQ4 How does the proposed method compare to state-of-the-art semi-supervised segmentation models in terms of segmentation accuracy and robustness on limited labeled data?
• RQ5 To what extent can the model reduce reliance on labeled data while maintaining high performance in real-world clinical applications?

Key findings

• UDC-Net achieves a Dice score of 77.4% ± 0.14, representing a 6.3% improvement over the fully supervised V-Net baseline (71.1%).
• The method outperforms other state-of-the-art semi-supervised models, including Mean Teacher, UA-MT, TCSM, and CCT, by at least 1.8% in Dice and 2.2% in Jaccard.
• With only 32 labeled scans and 140 unlabeled scans, UDC-Net achieves 75.0% Dice and 61.5% Jaccard, matching the performance of CCT with double the labeled data.
• When using 65 labeled scans and 708 unlabeled scans, UDC-Net achieves 77.4% Dice and 64.5% Jaccard, outperforming CCT with five times more unlabeled data.
• The model achieves a mean surface distance (ASD) of 3.9 mm, significantly lower than all compared methods, indicating superior contour accuracy.
• Ablation studies confirm that dual uncertainty (confidence + consensus) is more effective than single uncertainty, as it provides complementary guidance and improves robustness.