Introduction
Colorectal cancer remains one of the most prevalent and deadly cancers globally. While colonoscopy is the gold standard for early detection, its effectiveness is often limited by human factors, with approximately one in four polyps being missed during procedures. Automating polyp segmentation via computer vision presents a significant technical challenge due to "ambiguous boundaries." Polyps frequently exhibit colors and textures nearly identical to the surrounding healthy mucosa, leading to low contrast. Traditional Convolutional Neural Networks (CNNs) often struggle with these cases due to their limited receptive fields, which focus on local textures at the expense of global context, resulting in blurry or inaccurate edge segmentation.
Methods: The SegT Architecture
To address these challenges, the authors propose SegT, a Separated Edge-guidance Transformer network. Unlike standard CNN-based approaches, SegT leverages a Transformer architecture to capture long-range dependencies and global context.
The framework consists of three primary novel components:
- Transformer Encoder (PVT): Utilizing a Pyramid Vision Transformer (PVT) as the backbone, the model employs self-attention mechanisms to understand the broader global context of the colon wall.
- Separated Edge-guidance (SEG) Module: This core innovation mimics human perception. It includes a Separator (SE) Block that uses dual streams (foreground and background) to highlight boundaries, and an Edge-guidance (EG) Block that embeds specific edge information from an Edge Extractor Module (EEM) to sharpen boundary sensitivity.
- Cascade Fusion Module (CFM): This module progressively integrates semantic and positional information across multiple levels, refining the final segmentation map.
Experimental Results
SegT was evaluated on five major public datasets, including both "seen" data (Kvasir-SEG, CVC-ClinicDB) and "unseen" data (CVC-ColonDB, EndoScene, ETIS) to test generalization across different clinical environments.
Performance Metrics
SegT achieved state-of-the-art (SOTA) results across all key metrics (mDice, mIoU, and MAE):
- Accuracy: On the Kvasir-SEG dataset, SegT reached a mean Dice score (mDice) of 0.927, significantly outperforming U-Net (0.818) and specialized models like PraNet (0.898).
- Generalization: On the challenging, unseen CVC-ColonDB dataset, SegT outperformed the top CNN-based model (CaraNet) by 4.1% in mDice and 4.3% in mIoU.
- Qualitative Success: Visual results demonstrate that SegT excels at identifying "camouflaged" polyps and maintaining sharp boundaries even under difficult lighting conditions or occlusion.
Conclusion
SegT represents a significant advancement in polyp segmentation by combining the global modeling power of Transformers with a specialized edge-guidance mechanism. While the model shows exceptional accuracy and generalization, future work will focus on optimizing inference complexity for real-time clinical use and extending the architecture to 3D medical imaging to better capture volumetric data.
Summary updated