MEGANet-W: A Wavelet-Driven Edge-Guided Attention Framework for Weak Boundary Polyp Detection

Tan Zhe Yee, Ashwaq Qasem2025 · DOI

TL;DR

This paper introduces MEGANet-W, a novel deep learning framework for detecting colorectal polyps, particularly those with weak or low-contrast boundaries. By integrating a parameter-free Haar wavelet edge extraction module and a Wavelet Edge-Guided Attention (W-EGA) mechanism into the decoder of a U-Net architecture, MEGANet-W significantly improves polyp segmentation accuracy and boundary delineation without increasing model complexity.

What Problem Does This Paper Address?

Colorectal cancer (CRC) is a significant health concern, and early detection through colonoscopy is crucial. Polyps are key indicators, but their accurate segmentation from surrounding tissues is challenging due to subtle visual similarities in color and texture, especially when their boundaries are weak or low-contrast. Existing deep learning models often struggle with these fine edge details, either blurring them or relying on handcrafted features that are not robust to varying imaging conditions. This leads to missed detections and reduced diagnostic accuracy. The paper aims to address this by developing a more robust and precise polyp segmentation method that excels in challenging low-contrast scenarios.

Methods

MEGANet-W builds upon the U-Net architecture, incorporating a novel Wavelet-Driven Edge-Guided Attention (W-EGA) module. The core innovations are:

  • Two-Level Haar Wavelet Edge Head: Instead of traditional Laplacian-based edge detection, MEGANet-W employs a parameter-free, two-level Haar wavelet transform. This method decomposes features into low-frequency approximations and high-frequency detail sub-bands (LH, HL, HH) at different resolutions. This allows for the capture of both fine and coarser boundary cues without introducing learnable parameters, providing directional high-frequency information.
  • Wavelet Edge-Guided Attention (W-EGA) Modules: These modules are integrated into each decoder stage. They fuse three key information sources:
  • Reverse-Attention Branch: Similar to the original MEGANet, this branch uses inverted side-predictions to suppress confident foreground regions and highlight potential boundaries.
  • Boundary-Attention Branch: This branch leverages the wavelet-derived edge features from the decoder's current feature map to create a soft contour mask, emphasizing boundary regions.
  • Input-Edge Branch: This branch injects high-frequency priors from the original grayscale input image, using wavelet edges to provide context independent of the current logits.

These fused features are then processed through a CBAM (Convolutional Block Attention Module) to recalibrate feature responses, followed by a residual connection to preserve original semantics.

Datasets

The paper evaluates MEGANet-W on five standard benchmark datasets for polyp segmentation:

  • Kvasir-SEG
  • CVC-ClinicDB
  • CVC-ColonDB
  • ETIS-LaribPolypDB
  • EndoScene

For fair comparison, they used a subset of 1,450 images from Kvasir-SEG and CVC-ClinicDB, consistent with the original MEGANet paper.

Key Insights and Ideas

  • Wavelets for Edge Enhancement: The paper demonstrates that parameter-free Haar wavelets are a powerful tool for extracting rich, directional edge information without the need for learned parameters, offering a more robust alternative to traditional edge detectors.
  • Decoder-Stage Integration is Key: Injecting wavelet-derived edge features directly into the decoder's attention mechanism, rather than just the encoder, proves highly effective for refining boundary predictions in low-contrast regions.
  • Multi-Source Attention Fusion: Combining reverse attention, boundary attention (from wavelets), and input-edge information within the W-EGA module creates a synergistic effect, leading to more precise and reliable polyp segmentation.
  • Parameter Efficiency: The proposed wavelet head is parameter-free, and the W-EGA module introduces only a few learnable parameters (in the CBAM and attention gating), making the model computationally efficient and easier to train.

Key Results

MEGANet-W consistently outperforms seven state-of-the-art methods across the evaluated datasets.

  • Quantitative: It achieves the best or second-best performance in mIoU, mDice, and MAE. Notably, it sets new SOTA results on CVC-300 (mIoU 82.8%, mDice 89.5%), Kvasir-SEG (mIoU 87.0%, mDice 92.0%), and ETIS (mIoU 69.6%, mDice 77.5%). It also delivers the best MAE values, indicating improved pixel-level precision and boundary localization.
  • Qualitative: Visual comparisons show MEGANet-W providing more precise boundary delineation in low-contrast cases, better handling of polyps with similar color/texture to the background, cleaner segmentations in the presence of illumination artifacts, and accurate segmentation of small, spatially separated polyps without merging errors.

Limitations and Future Work

  • Limitations: The paper acknowledges that while MEGANet-W shows strong performance, further validation on diverse, multi-center clinical datasets is needed.
  • Future Work: The authors suggest extending the work to video colonoscopy or volumetric data. They also highlight the importance of enhancing interpretability through visualization of wavelet attention responses to build clinical trust and facilitate deployment.

Summary updated

BibTeX
@misc{yee2025meganet,
  author = {Tan Zhe Yee and Ashwaq Qasem},
  title = {{MEGANet-W: A Wavelet-Driven Edge-Guided Attention Framework for Weak Boundary Polyp Detection}},
  year = {2025},
  doi = {10.1109/iceei68459.2025.11331219},
}
Figure