Introduction
Colorectal cancer (CRC) is a leading cause of cancer-related mortality worldwide. Early detection through polyp segmentation in colonoscopy images is a critical step in prevention and treatment. While deep learning has significantly advanced this field, most state-of-the-art models rely on fully supervised learning, requiring vast amounts of pixel-level manual annotations by medical experts. Such annotation is time-consuming, expensive, and subjective.
To address this, the authors propose a semi-supervised framework that effectively leverages a small amount of labeled data alongside a larger pool of unlabeled images. By utilizing collaborative networks and a dual-adversarial training strategy, the model achieves high accuracy while significantly reducing the need for manual labeling.
Methodology
The proposed framework is built upon a "Collaborative and Adversarial Learning" architecture designed to handle the high variability in polyp shape, size, and texture.
Collaborative Representation Learning
Instead of a single segmentation backbone, the framework employs two collaborative modules to capture different aspects of polyp morphology:
- Focused Extraction Module (FEM): This module uses global attention and average pooling to zero in on the polyp’s primary location and spatial features.
- Dispersive Extraction Module (DEM): This module utilizes dilated convolutions at varying rates to capture "scattered" boundary information, ensuring that the edges of the polyp are precisely defined.
A mutual consistency constraint is applied between these two networks, forcing them to "agree" on their predictions. This collaborative approach reduces the individual biases that often arise when training on limited datasets.
Dual Adversarial Learning
The framework incorporates two discriminators to maximize the utility of unlabeled data:
- Primary Discriminator ($D_1$): This network produces "confidence maps" for unlabeled predictions. These maps identify high-certainty regions, effectively turning noisy unlabeled data into reliable "pseudo-labels" for the segmentation networks.
- Auxiliary Adversarial Learning (AAL): A second discriminator ($D_2$) is used to distinguish between predictions from labeled versus unlabeled images. This encourages the model to produce high-quality results for unlabeled data that are indistinguishable from those backed by ground truth.
Experimental Results
The model was evaluated on two major benchmarks: Kvasir-SEG and CVC-Clinic DB. The results demonstrate that the proposed semi-supervised approach is highly competitive with, and sometimes superior to, fully supervised methods.
Key Performance Metrics
- Data Efficiency: On the Kvasir-SEG dataset, using only 30% of the labeled data, the model achieved a Dice coefficient of 0.8095.
- Competitive Edge: Remarkably, this semi-supervised result outperformed several fully supervised models (using 100% labels), including ResU-Net (0.7647) and U-Net++ (0.8045).
- Robustness: The model showed a significant lead in Mean Absolute Error (MAE), indicating fewer large-scale errors in polyp estimation compared to existing semi-supervised methods like Hung’s method.
Conclusion
This research presents a robust semi-supervised framework for polyp segmentation that balances global localization with local boundary precision. By integrating collaborative feature extraction with dual-adversarial learning, the model achieves state-of-the-art performance with minimal labeled data. Future work will focus on extending this framework to video data to leverage temporal consistency for real-time clinical applications.
Summary updated