Polyp Segmentation Using Wavelet-Based Cross-Band Integration for Enhanced Boundary Representation
Haesung Oh, Jaesung Lee · 2026
Read Polyp Segmentation Using Wavelet-Based Cross-Band Integration for Enhanced Boundary Representation by Haesung Oh, Jaesung Lee (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2603-03682.txt.
Introduction
Haesung Oh and Jaesung Lee outline colon polyps more precisely by listening to grayscale. Polyp segmentation guides early colorectal cancer detection, but boundaries blur under low contrast, uneven lighting, and tissue that mimics polyp color. Color-based models inherit all three problems. Their 2026 paper first asks a basic question in the wavelet domain: where does boundary contrast actually live?
Across every frequency band, grayscale representations preserve stronger polyp-to-background contrast than full color images. Boundary cues, in other words, are cleaner without color. The proposed model therefore fuses grayscale and color streams through frequency-consistent interaction, sharpening edges while keeping the overall structure coherent. Extensive experiments on four benchmark datasets show superior boundary precision and robustness over conventional models, especially where color alone misleads.
Catalog record
| CREATOR | Haesung Oh, Jaesung Lee |
|---|---|
| YEAR | 2026 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 9731 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2603.03682
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
Polyp Segmentation Using Wavelet-Based Cross-Band Integration for Enhanced Boundary Representation Haesung Oh, Jaesung Lee Abstract Accurate polyp segmentation is essential for early colorectal cancer detection, yet achieving reliable boundary localization remains challenging due to low mucosal contrast, uneven illumination, and color similarity between polyps and surrounding tissue. Conventional methods relying solely on RGB information often struggle to delineate precise boundaries due to weak contrast and ambiguous structures between polyps and surrounding mucosa. To establish a quantitative foundation for this limitation, We analyzed polyp–background contrast in the wavelet domain, revealing that grayscale representations consistently preserve higher boundary contrast than RGB images across all frequency bands. This finding suggests that boundary cues are more distinctly represented in the grayscale domain than in the color domain. Motivated by this finding, we propose a segmentation model that integrates grayscale and RGB representations through complementary frequency-consistent interaction, enhancing boundary precision while preserving structural coherence. Extensive experiments on four benchmark datasets demonstrate that the proposed approach achieves superior boundary precision and robustness compared to conventional models. 1 Introduction Accurate polyp segmentation is vital for early colorectal cancer detection but remains difficult due to low mucosal contrast, uneven illumination, and strong visual similarity between polyps and surrounding mucosa 1 . These conditions blur polyp boundaries, particularly in small or flat cases, making reliable delineation challenging. Although recent boundary-aware methods improve edge perception, their dependence on RGB inputs limits robustness under contrast variations. To investigate this limitation, we performed a wavelet-based contrast analysis between RGB and grayscale images. The contrast index (CI), defined as \text{CI}=|\mu_{\text{polyp}}-\mu_{\text{background}}|/(\mu_{\text{polyp}}+\mu_{\text{background}}+\epsilon) , measures the distinction between polyp and background regions using the mean of absolute wavelet coefficients. As shown in Fig. 1 , grayscale consistently achieves higher CI across all sub-bands, indicating that boundary cues are more distinct in the intensity domain. Building on this evidence, we propose a segmentation approach that integrates grayscale and RGB representations through frequency-band interaction. By encouraging interaction between corresponding wavelet sub-bands of both modalities, contrast-rich grayscale features refine RGB-derived spatial structures, improving boundary precision while preserving overall coherence. 2 Related Work Polyp segmentation has become a key task in computer-aided colorectal cancer screening, where precise boundary delineation is essential for accurate diagnosis and treatment planning. A variety of deep learning–based models have been developed for polyp segmentation 11 ; 15 ; 14 ; 8 ; 5 . Among these, boundary-aware models such as PraNet 6 , CaraNet 10 , MEGANet 4 , and Polyper 12 aim to improve boundary localization through attention- and edge-guided mechanisms for more accurate polyp delineation. Nevertheless, most of these approaches primarily rely on RGB representations, which capture chromatic appearance but insufficiently describe structural contrast. Figure 1: Structural contrast comparison between RGB and grayscale images in the wavelet domain, showing consistently higher contrast for grayscale across all detail sub-bands. Figure 2: Proposed wavelet-based cross-band integration framework that fuses frequency-consistent information from RGB and grayscale features for enhanced boundary representation. 3 Method The proposed model adopts a dual-encoder structure designed to leverage the complementary characteristics of RGB and grayscale modalities. Each encoder is based on Res2Net 7 and extracts hierarchical feature representations: the RGB encoder captures chromatic and textural cues, while the grayscale encoder focuses on contrast-driven structural patterns that are effective for boundary discrimination. Features from corresponding encoder stages are processed within the decoder through two key components: the Band-Specific Window Cross-Attention (BS-WCA) module and the Cascade Dilated Fusion (CDF) block. The BS-WCA module performs frequency-aligned interaction between the RGB and grayscale features by selectively exchanging information across identical wavelet sub-bands, allowing high-frequency grayscale details to refine RGB-derived structural features. The CDF block then integrates the refined multi-scale features through dilated convolutions, preserving both fine-grained boundary precision and global contextual consistency. As illustrated in Fig. 2 , the integration of BS-WCA and CDF within a dual-encoder–decoder framework enables model to exploit contrast-rich intensity information while maintaining the contextual advantages of RGB representations, resulting in more stable and accurate polyp boundary segmentation. 4 Experimental Results Datasets and Evaluation The experiments utilized four widely used polyp segmentation datasets: Kvasir-SEG (1,000 images) 9 , CVC-ClinicDB (612 images) 3 , CVC-ColonDB (380 images) 2 , and ETIS (196 images) 13 . All datasets were partitioned into training, validation, and testing sets to ensure consistent evaluation across varying data distributions. Performance was measured using the Dice coefficient and Intersection over Union (IoU), following standard practices in polyp segmentation. Implementation Details All experiments were implemented using PyTorch and Torchvision and conducted on a single NVIDIA GPU (CUDA 11.7, Ubuntu 20.04, Python 3.9, PyTorch 1.13, Torchvision 0.14). The proposed model and all baseline models were trained and evaluated with a batch size of 8. Results. As shown in Table 1 , the proposed method consistently outperforms existing approaches across all four benchmark datasets in terms of both Dice and IoU scores. The improvement is particularly evident in the overall average performance, demonstrating the benefit of jointly leveraging grayscale and RGB representations. Compared to recent conventional models, the proposed model achieves higher mean Dice and IoU values while maintaining stable performance across datasets of varying scales and imaging conditions. These results indicate that the integration of intensity-based and color-based representations provides a more comprehensive feature space for polyp segmentation. Table 1: Quantitative comparison of polyp segmentation methods across four benchmark datasets, evaluated by mean Dice and IoU scores (mean ± standard deviation) averaged over 10 random runs. Methods Kvasir ClinicDB ColonDB ETIS mDice mIoU mDice mIoU mDice mIoU mDice mIoU Ours 0.885 ± 0.021 0.822 ± 0.019 0.926 ± 0.014 0.862 ± 0.023 0.913 ± 0.021 0.840 ± 0.042 0.922 ± 0.029 0.821 ± 0.029 Polyper 0.867 ± 0.014 0.796 ± 0.020 0.914 ± 0.019 0.841 ± 0.021 0.868 ± 0.042 0.796 ± 0.035 0.888 ± 0.047 0.760 ± 0.047 MEGANet 0.863 ± 0.011 0.802 ± 0.018 0.909 ± 0.011 0.801 ± 0.077 0.802 ± 0.083 0.704 ± 0.059 0.747 ± 0.097 0.598 ± 0.077 CRCNet 0.879 ± 0.015 0.815 ± 0.018 0.910 ± 0.052 0.854 ± 0.018 0.866 ± 0.064 0.800 ± 0.046 0.665 ± 0.027 0.582 ± 0.129 CaraNet 0.727 ± 0.023 0.628 ± 0.035 0.836 ± 0.006 0.702 ± 0.014 0.760 ± 0.056 0.651 ± 0.038 0.784 ± 0.078 0.633 ± 0.105 ConvSegNet 0.856 ± 0.008 0.765 ± 0.022 0.902 ± 0.020 0.810 ± 0.025 0.884 ± 0.033 0.756 ± 0.044 0.859 ± 0.055 0.667 ± 0.077 DUCKNet 0.818 ± 0.016 0.751 ± 0.019 0.878 ± 0.026 0.791 ± 0.033 0.683 ± 0.161 0.570 ± 0.073 0.383 ± 0.205 0.321 ± 0.099 PraNet 0.650 ± 0.021 0.524 ± 0.032 0.793 ± 0.032 0.648 ± 0.030 0.784 ± 0.063 0.620 ± 0.037 0.666 ± 0.090 0.423 ± 0.092 UNet 0.775 ± 0.013 0.668 ± 0.025 0.855 ± 0.019 0.762 ± 0.024 0.802 ± 0.083 0.704 ± 0.059 0.549 ± 0.186 0.382 ± 0.099 5 Discussion This study primarily evaluates region-level segmentation performance using Dice and IoU. To better understand the behavior of the proposed model, further analyses should include boundary-sensitive metrics for contour accuracy, qualitative comparisons across imaging conditions, and frequency-domain ablations to assess modality contributions. Such evaluations would provide a more comprehensive view of how grayscale–RGB integration influences segmentation performance. 6 Conclusion This work introduced a dual-encoder segmentation framework