[Paper Review] Semi-supervised Left Atrium Segmentation with Mutual Consistency Training

Yicheng Wu, Minfeng Xu+3 more|arXiv (Cornell University)|Mar 4, 2021

Cardiac Valve Diseases and Treatments18 references20 citations

TL;DR

This paper proposes MC-Net, a semi-supervised 3D left atrium segmentation method that enhances model generalization by leveraging unlabeled challenging regions through mutual consistency training. By using two slightly different decoders and a cycled pseudo label scheme to enforce low-entropy, consistent predictions, MC-Net achieves state-of-the-art performance, outperforming six SOTA methods on the LA database with 20% labeled data.

ABSTRACT

Semi-supervised learning has attracted great attention in the field of machine learning, especially for medical image segmentation tasks, since it alleviates the heavy burden of collecting abundant densely annotated data for training. However, most of existing methods underestimate the importance of challenging regions (e.g. small branches or blurred edges) during training. We believe that these unlabeled regions may contain more crucial information to minimize the uncertainty prediction for the model and should be emphasized in the training process. Therefore, in this paper, we propose a novel Mutual Consistency Network (MC-Net) for semi-supervised left atrium segmentation from 3D MR images. Particularly, our MC-Net consists of one encoder and two slightly different decoders, and the prediction discrepancies of two decoders are transformed as an unsupervised loss by our designed cycled pseudo label scheme to encourage mutual consistency. Such mutual consistency encourages the two decoders to have consistent and low-entropy predictions and enables the model to gradually capture generalized features from these unlabeled challenging regions. We evaluate our MC-Net on the public Left Atrium (LA) database and it obtains impressive performance gains by exploiting the unlabeled data effectively. Our MC-Net outperforms six recent semi-supervised methods for left atrium segmentation, and sets the new state-of-the-art performance on the LA database.

Motivation & Objective

• To address the challenge of limited annotated medical data in left atrium segmentation by effectively utilizing unlabeled data.
• To improve model generalization by emphasizing difficult, ambiguous regions (e.g., small branches, blurred edges) in the absence of labels.
• To reduce model uncertainty in challenging regions by leveraging epistemic uncertainty as a supervisory signal.
• To develop a semi-supervised framework that outperforms existing methods on the public LA database.
• To demonstrate that mutual consistency between decoders enhances feature learning from unlabeled, high-uncertainty regions.

Proposed method

• MC-Net employs a shared encoder and two slightly different decoders to generate diverse yet consistent predictions for the same input.
• Prediction discrepancies between the two decoders are used to estimate uncertainty and guide training.
• A cycled pseudo label scheme transforms prediction disagreements into an unsupervised loss to encourage mutual consistency.
• Soft pseudo labels are generated via a sharpening function and used with entropy regularization to minimize prediction uncertainty.
• The method applies consistency regularization by minimizing the discrepancy between the two decoders’ outputs, promoting low-entropy, stable predictions.
• The framework is trained end-to-end using both labeled data and unlabeled data, with the cycled pseudo label mechanism dynamically refining predictions.

Experimental results

Research questions

• RQ1 Can mutual consistency training between two decoders improve generalization in semi-supervised left atrium segmentation?
• RQ2 Can unlabeled challenging regions—previously underestimated—be effectively leveraged to reduce model uncertainty?
• RQ3 Does a cycled pseudo label scheme that enforces consistency between decoders lead to better segmentation performance than standard consistency regularization?
• RQ4 Can MC-Net achieve state-of-the-art performance using only 20% labeled data on the LA database?
• RQ5 How does model diversity (via different decoders) impact performance compared to identical decoders?

Key findings

• MC-Net achieves a Dice score of 90.34% on the LA database using only 20% labeled data, outperforming the fully supervised V-Net baseline (91.14% with 100% labels).
• With 20% labeled data, MC-Net achieves a 95% Hausdorff Distance (95HD) of 6.00 voxels and an average surface distance (ASD) of 1.77 voxels, significantly outperforming all six SOTA methods.
• The ablation study confirms that using two different decoders (V2d-Net) improves performance over identical decoders (V2-Net), with a Dice gain of 1.18% under 10% labeled data.
• Adding the cycled pseudo label (CPL) component reduces 95HD from 9.36 to 6.00 voxels and ASD from 2.18 to 1.77 voxels under 20% labeled data, demonstrating its effectiveness.
• The model produces more complete left atrium segmentation with fewer isolated regions, especially in challenging areas (e.g., small branches), without post-processing.
• The method is compatible with other shape-constrained models and can be extended to improve performance in other semi-supervised medical image segmentation tasks.