Abstract
Semi-supervised learning (SSL) is a critical paradigm in medical image segmentation, aiming to leverage large amounts of unlabeled data to mitigate the scarcity of expert annotations. However, traditional Mean Teacher (MT) frameworks often suffer from "cognitive bias," where the student and teacher networks become overly similar, reinforcing prediction errors—especially at challenging tissue boundaries. We propose the Mutual Correction Framework (MCF), a dual-subnet architecture designed to break this cycle. MCF utilizes heterogeneous subnets to maintain diversity and introduces two novel modules: Contrastive Difference Review (CDR) for supervised error rectification and Dynamic Competitive Pseudo-Label Generation (DCPLG) for adaptive unlabeled learning. Our framework achieves state-of-the-art performance on MRI and CT benchmarks, significantly improving boundary accuracy.
Introduction
In medical imaging, pixel-level annotations are labor-intensive and require specialized expertise. Semi-supervised learning addresses this by training models on a small set of labeled data and a larger set of unlabeled data. A common approach is the Mean Teacher framework, which enforces consistency between a student network and a teacher network (a moving average of the student).
Despite its popularity, MT frameworks are prone to cognitive bias. Because the teacher is derived from the student, they often share the same "blind spots." When the model misclassifies pixels at organ boundaries, it tends to reinforce these mistakes over time, lacking a mechanism for self-correction. MCF addresses this by replacing the student-teacher mirror with a competitive dual-subnet system that identifies and corrects discrepancies.
Methodology
Heterogeneous Dual-Subnet Architecture
To prevent the networks from converging to the same errors, MCF employs two different architectures: a standard VNet and a 3D-ResVNet. This structural diversity ensures that the subnets develop different feature representations, making their disagreements a meaningful signal for potential errors.
Contrastive Difference Review (CDR)
The CDR module operates on labeled data. It identifies "discrepancy regions" where the two subnets disagree using an XOR operation. By focusing on these specific pixels, the framework applies a rectification loss ($L_{rec}$), forcing both networks to review and correct their predictions against the ground truth. This targeted learning effectively reduces cognitive bias at the most difficult-to-segment regions.
Dynamic Competitive Pseudo-Label Generation (DCPLG)
For unlabeled data, MCF implements a dynamic teacher selection process. Rather than a fixed teacher, the framework evaluates both subnets in real-time using the Dice loss calculated on the current batch of labeled data. The subnet with the higher performance (the "winner") generates pseudo-labels to guide the other subnet (the "loser"). This ensures that the highest quality supervision is always used to train the models on unlabeled volumes.
Experimental Results
Quantitative Performance
MCF was evaluated on the Left Atrial (LA) MRI and NIH Pancreas CT datasets. On the LA dataset with 20% labeled data, MCF achieved a Dice score of 88.71%, surpassing previous state-of-the-art (SOTA) methods like UA-MT (87.65%).
Notably, MCF showed superior performance in boundary-sensitive metrics. The 95% Hausdorff Distance (95HD) on the LA dataset dropped from 9.70 (SOTA) to 6.32, indicating much more precise edge detection. Similar gains were observed on the NIH Pancreas dataset, where MCF reached a Dice score of 75.00%.
Qualitative Analysis
Visual comparisons demonstrate that MCF produces cleaner segmentations with fewer isolated false positives and smoother boundaries compared to traditional MT-based approaches.
Conclusion
The Mutual Correction Framework effectively mitigates cognitive bias in semi-supervised medical image segmentation. By leveraging architectural diversity and dynamic competition, MCF identifies and rectifies errors that traditional consistency-based methods overlook. The significant reduction in boundary errors makes MCF a robust candidate for clinical applications where precision is paramount.
Summary updated