Introduction
Deep learning has significantly advanced medical image segmentation, yet the reliance on extensive pixel-level annotations remains a major hurdle. Expert clinicians' time is costly, making large-scale labeled datasets rare. Semi-supervised learning (SSL) offers a solution by training models on a small amount of labeled data and a large pool of unlabeled data.
However, traditional SSL methods, such as the Mean-Teacher (MT) framework, often suffer from "cognitive bias." In MT, the teacher is a temporal average of the student; if the student makes an error, the teacher reinforces it, creating an echo chamber that hinders performance. UC-Seg addresses this by introducing a cross-training framework that utilizes structural diversity and explicit uncertainty modeling.
The UC-Seg Framework
UC-Seg moves away from the traditional student-teacher paradigm, instead employing two structurally different subnets—a U-Net and a V-Net—to perform mutual correction.
Core Components
- Cross-subnet Consistency Preservation (CCP): This module ensures that both subnets align their feature representations. It includes Intra-subnet Feature Enhancement (IFE), which uses contrastive learning and a "feature bank" to cluster features of the same class together in the embedding space while pushing different classes apart.
- Uncertainty-aware Pseudo-label Generation (UPG): This is the core of the pseudo-labeling process. Instead of simple thresholding, UPG uses Subjective Logic (via a Dirichlet distribution) to estimate pixel-wise uncertainty. A cross-attention mechanism fuses predictions from both subnets, giving higher weight to confident regions and lower weight to uncertain ones, resulting in cleaner training signals.
Experimental Results
The authors evaluated UC-Seg across seven diverse tasks, including five colonoscopy datasets, skin lesion detection, and various 3D volumetric datasets (Brain, Cardiac, and Pancreas).
UC-Seg consistently outperformed state-of-the-art (SOTA) methods like BCP and CAML. For instance, on the CVC-ClinicDB polyp segmentation task with only 10% labeled data, UC-Seg achieved a Dice score of 80.63%, significantly higher than its closest competitors. Qualitative results demonstrate that UC-Seg produces segmentation maps that are much closer to the ground truth, especially in challenging regions where other models fail.
Conclusion and Future Work
UC-Seg effectively mitigates cognitive bias in semi-supervised medical image segmentation by leveraging architectural diversity and uncertainty-aware fusion. The model's ability to generalize to unseen datasets highlights its robustness. While the dual-subnet architecture increases computational costs during training, the gains in accuracy and reliability make it a promising framework for real-world clinical applications where labeled data is scarce. Future research will explore refining the UPG module using unlabeled data to further enhance its performance.
Summary updated