Abstract
Accurate segmentation of polyps in colonoscopy images is essential for early-stage diagnosis and management of colorectal cancer. Despite advancements in deep learning for polyp segmentation, enduring limitations persist. The edges of polyps are typically ambiguous, making them difficult to discern from the background, and the model performance is often compromised by the influence of irrelevant or unimportant features. To alleviate these challenges, we propose a novel model named Edge-Prioritized Polyp Segmentation (EPPS). Specifically, we incorporate an Edge Mapping Engine (EME) aimed at accurately extracting the edges of polyps. Subsequently, an Edge Information Injector (EII) is devised to augment the mask prediction by injecting the captured edge information into Decoder blocks. Furthermore, we introduce a component called Selective Feature Decoupler (SFD) to suppress the influence of noise and extraneous features on the model. Extensive experiments on 3 widely used polyp segmentation benchmarks demonstrate the superior performance of our method compared with other state-of-the-art approaches.
Introduction
Colorectal cancer prevention relies heavily on the early detection and removal of polyps during colonoscopy. Automated polyp segmentation in video feeds can assist clinicians, but it faces two primary hurdles: boundary ambiguity and environmental noise. Polyps often share visual characteristics with the surrounding intestinal wall, resulting in blurry edges. Additionally, colonoscopy environments are prone to specular reflections (glare), fluid, and inconsistent lighting, which act as distractors for standard AI models. Existing models often struggle to separate the "signal" (the polyp) from the "noise" (the background clutter).
Methodology
The Edge-Prioritized Polyp Segmentation (EPPS) model addresses these challenges through a specialized architecture built upon a ResNet-50 backbone enhanced by three novel components.
Edge Mapping and Injection
The Edge Mapping Engine (EME) is dedicated to boundary extraction. It utilizes multi-level feature processing to generate a precise edge map. This edge information is then fed into the Edge Information Injector (EII), which integrates it into the decoder blocks. By employing Channel and Spatial Attention Modules (CAM/SAM), the EII ensures the model maintains focus on polyp boundaries throughout the upsampling process.
Selective Feature Decoupling
To mitigate the impact of environmental noise, the Selective Feature Decoupler (SFD) is introduced between the encoder and decoder. The SFD utilizes Mutual Information Neural Estimation (MINE) to partition feature maps into "significant" (polyp-related) and "unimportant" (noise-related) components. By minimizing the mutual information between these sets, the model effectively filters out irrelevant data, such as camera glare, before it reaches the decoder.
Experiments and Results
The EPPS model was evaluated on three widely recognized benchmarks: Kvasir-SEG, CVC-ClinicDB, and Kvasir-Sessile.
Quantitative Performance
EPPS demonstrated superior performance across all datasets. On Kvasir-SEG, it achieved a mean Dice Similarity Coefficient (mDSC) of 0.930, a 2.3% improvement over the previous best method. Notably, on the Kvasir-Sessile dataset—which contains challenging flat lesions—EPPS showed a significant improvement of 3.2% in mIoU over the second-best model, proving its effectiveness at finding subtle lesions.
Qualitative Analysis
Visual results confirm that EPPS is more robust to environmental distractors. The SFD module successfully identifies and suppresses specular reflections that typically cause other models to produce false positive fragments in the mask.
Discussion and Conclusion
EPPS advances polyp segmentation by prioritizing edge information and explicitly decoupling noise from significant features. While the model achieves state-of-the-art accuracy, future work should address the computational overhead introduced by the MINE network and attention modules to ensure real-time performance in clinical settings. Furthermore, the model's reliance on traditional operators like Canny for generating training labels suggests potential for further refinement in edge definition. Validating the model on raw, unedited colonoscopy video streams remains a necessary step toward practical clinical deployment.
Summary updated