Abstract
Semi-supervised learning (SSL) is a critical approach in medical image segmentation, where high-quality expert annotations are often scarce. We present the Mutual Consistency Network (MC-Net), a novel framework designed to leverage unlabeled data by focusing on high-uncertainty "challenging regions." MC-Net employs a dual-decoder architecture with a Cycled Pseudo Label (CPL) scheme to encourage consistency between diverse up-sampling strategies. Evaluated on the 2018 Atrial Segmentation Challenge, MC-Net achieves state-of-the-art performance, reaching a 90.34% Dice score with only 20% labeled data, nearly matching the performance of fully supervised models.
Introduction
Deep learning models for 3D medical image segmentation typically require dense, voxel-by-voxel annotations, which are labor-intensive and require specialized expertise. Semi-supervised learning aims to mitigate this by training on a small labeled dataset alongside a larger pool of unlabeled data.
However, existing SSL methods often struggle with "challenging regions"—areas characterized by blurred edges or intricate anatomical structures. While many models provide unreliable guidance in these high-uncertainty areas, we argue that these regions contain the most crucial information for improving model generalization. MC-Net is designed specifically to exploit these difficult, unlabeled regions to boost segmentation accuracy.
Methods
The Mutual Consistency Network (MC-Net) introduces three primary innovations to the standard segmentation pipeline:
- Dual-Decoder Architecture: MC-Net utilizes a single shared encoder with two architecturally distinct decoders. Decoder A employs transposed convolutions for up-sampling, while Decoder B uses tri-linear interpolation. This architectural diversity ensures that the decoders produce slightly different outputs, capturing "epistemic uncertainty" (the model's own lack of knowledge) in a single forward pass.
- Cycled Pseudo Label (CPL) Scheme: To enforce mutual consistency, the model generates "soft pseudo labels" from one decoder's output to supervise the other. These labels are processed through a sharpening function to reduce entropy. This cross-teaching mechanism encourages both decoders to reach a consensus, particularly in high-uncertainty regions.
- Composite Loss Function: Training is guided by a supervised Dice loss for labeled data and an unsupervised consistency loss (Mean Squared Error) between a decoder's prediction and the other's pseudo label for all data.
Experiments and Results
The model was primarily evaluated on the 2018 Atrial Segmentation Challenge (LA database), consisting of 100 3D MRI scans. We tested the framework under 10% (8 scans) and 20% (16 scans) labeled data settings.
MC-Net established a new state-of-the-art (SOTA) on the LA database, outperforming recent methods such as UA-MT, SASSNet, and DTC. With only 20% labeled data, MC-Net achieved a 90.34% Dice score, remarkably close to the 91.14% upper bound achieved by a fully supervised V-Net. In the 10% setting, the model reached an 87.71% Dice score, significantly outperforming the baseline V-Net (79.99%).
Qualitative analysis indicates that MC-Net produces cleaner and more anatomically complete segmentations, particularly in the complex "branches" of the left atrium, where other methods often produce fragmented results.
Conclusion
MC-Net demonstrates that uncertainty in medical imaging can be treated as a powerful supervision signal rather than mere noise. By utilizing architectural diversity and the Cycled Pseudo Label scheme, the model effectively learns robust features from challenging, unlabeled regions. While the dual-decoder setup increases parameter count, its efficiency over traditional Monte Carlo uncertainty estimation makes it a viable candidate for high-performance, semi-supervised medical image analysis.
Summary updated