Ma mba G uided Bo u ndary P rior Matters: A New Perspective for Generalized Polyp Segmentation
Tapas K. Dutta, Snehashis Majhi, Deepak Ranjan Nayak, Debesh Jha · 2025
Read Ma mba G uided Bo u ndary P rior Matters: A New Perspective for Generalized Polyp Segmentation by Tapas K. Dutta, Snehashis Majhi, Deepak Ranjan Nayak, Debesh Jha (2025) 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-2507-01509.txt.
Introduction
Tapas Dutta and co-authors sharpen polyp segmentation in colonoscopy images by teaching the model to respect boundaries. Their 2025 paper tackles a screening task where misses matter: polyps vary in shape, size, and color and often blend into the gut wall, so weak edges fool standard networks. Their design pairs a Mamba backbone, a fast sequence model, with a boundary prior that guides attention to the polyp's rim.
The aim is steadier outlines on the blurry cases and better transfer to new hospitals' images, both needed for real-time clinical use. The authors frame boundary awareness, rather than a bigger network, as the missing ingredient in this corner of cancer screening.
Catalog record
| CREATOR | Tapas K. Dutta, Snehashis Majhi, Deepak Ranjan Nayak, Debesh Jha |
|---|---|
| YEAR | 2025 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 24 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2507-01509
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
Ma mba G uided Bo u ndary P rior Matters: A New Perspective for Generalized Polyp Segmentation Tapas K. Dutta, Snehashis Majhi, Deepak Ranjan Nayak, Debesh Jha Abstract Polyp segmentation in colonoscopy images is crucial for early detection and diagnosis of colorectal cancer. However, this task remains a significant challenge due to the substantial variations in polyp shape, size, and color, as well as the high similarity between polyps and surrounding tissues, often compounded by indistinct boundaries. While existing encoder-decoder CNN and transformer-based approaches have shown promising results, they struggle with stable segmentation performance on polyps with weak or blurry boundaries. These methods exhibit limited abilities to distinguish between polyps and non-polyps and capture essential boundary cues. Moreover, their generalizability still falls short of meeting the demands of real-time clinical applications. To address these limitations, we propose SAM-MaGuP , a groundbreaking approach for robust polyp segmentation. By incorporating a boundary distillation module and a 1D-2D Mamba adapter within the Segment Anything Model (SAM), SAM-MaGuP excels at resolving weak boundary challenges and amplifies feature learning through enriched global contextual interactions. Extensive evaluations across five diverse datasets reveal that SAM-MaGuP outperforms state-of-the-art methods, achieving unmatched segmentation accuracy and robustness. Our key innovations—a Mamba-guided boundary prior and a 1D-2D Mamba block—set a new benchmark in the field, pushing the boundaries of polyp segmentation to new heights. 1 Introduction Colorectal Cancer (CRC) is one of the leading causes of cancer-related death across the globe, ranking among the third most frequently diagnosed cancer [ 3 , 18 ] . It starts from certain growths or polyps in the inner lining of the colon and often leads to CRC if not treated timely. The incidence of CRC can be reduced significantly with early diagnosis and on-time treatment, saving many lives. Colonoscopy is the gold standard screening procedure for identifying and removing polyps. However, this detection procedure is manual, time-consuming, and highly reliant on the clinician’s skill and expertise. Additionally, the miss rate of polyp detection is considerably high (6-27%) [ 1 ] . Therefore, it is of utmost significance to design automated polyp segmentation approaches to detect the polyp and provide clinicians with precise location and boundary information, reducing the miss rate. There has been remarkable progress in the development of polyp segmentation methods based on deep learning in the past years, and the results shown have been awe-inspiring [ 17 , 28 , 9 ] . In early efforts, U-Net and its variants, such as UNet++ and ResUNet++ [ 12 ] , have been adopted as a model of choice. However, these models fall short of capturing boundary details. Following these works, various notable architectures such as FCN [ 10 ] , PraNet [ 9 ] , MSNet [ 25 ] , CFA-Net [ 27 ] , and MEGANet [ 4 ] have been proposed. While such architectures exhibit their capability to capture the boundary cues and handle variability in the size of polyps, they encounter challenges in extracting global contextual information, which is crucial for polyp identification. Later, inspired by the remarkable success of transformers, a few architectures such as PVT-Cascade [ 16 ] and CTNet [ 23 ] has recently been introduced, which allows modeling global feature dependencies and learning prominent features and have demonstrated impressive segmentation results. Despite the significant progress, segmenting polyps has remained a daunting challenge due to the camouflage characteristics of polyps, i.e., intrinsic resemblance among polyps and the surrounding mucosa, considerable variability in polyp’s shape and size, and absence of sharp boundaries. This results in unstable segmentation performance and poor localization of polyps. In addition, the existing models have shown limited capacity to learn distinguishable features and detailed boundary cues from polyp regions. Segment Anything Model (SAM), a foundational model, has recently garnered the interest of researchers for segmenting images with exceptional generalization capabilities [ 13 ] . However, its performance is limited when applied to medical images because of inadequate domain-specific knowledge [ 26 ] . Additionally, it leads to high computational overhead and memory requirements [ 22 ] . To address these issues, an adapter module integrated SAM architecture has recently been proposed by leveraging Mamba [ 6 ] , an efficient approach known for its excellent capability to model long-range interactions with linear computational complexity [ 29 ] . Although this approach has obtained superior generalization performance with exceptional computational benefits, it does not take into account boundary information, leading to suboptimal polyp segmentation performance. Additionally, the generalizability needs to be improved further to meet real-time clinical requirements. To address the persistent challenges in polyp segmentation, we introduce SAM-MaGuP , a novel framework that leverages boundary distillation and the Mamba adapter to enhance SAM’s performance. The key building block MaGuP module revolutionizes general-adapter-based [ 5 ] fine-tuning by refining SAM’s feature representations to capture subtle boundary details often overlooked by conventional models. The module is composed of two transformative components, 1D-2D Mamba: A unique feature fusion technique that merges channel (1D) and spatial (2D) information, empowering SAM to learn global contexts at multiple scales, unlocking richer feature representations. Boundary Distillation Component (BDC): A dynamic unit that sharpens the model’s ability to detect exact boundary locations, even in ambiguous or weakly defined regions. Extensive experiments across five diverse datasets demonstrate that SAM-MaGuP effectively tackles blurry boundary issues, setting a new standard by outshining current state-of-the-art (SoTA) methods. In summary, our major contributions are: • We introduce a Mamba-guided boundary prior in SAM, the first of its kind, to effectively tackle the weak-boundary challenge in polyp segmentation. • We propose a boundary distillation component that empowers SAM to accurately detect polyps in weak-boundary scenarios, enhanced by a 1D-2D Mamba block that optimizes feature interactions across spatial and channel dimensions. • We demonstrate the outstanding performance of SAM-MaGuP on five diverse datasets, consistently surpassing existing SoTA polyp segmentation techniques. 2 Proposed SAM-MaGuP In this section, we present the SAM-MaGuP framework, specifically tailored to tackle the challenging task of weak-boundary polyp segmentation. By leveraging the versatility of the SAM backbone, we elevate its capabilities to address complex segmentation tasks. We first introduce SAM, a powerful model designed for general-purpose object segmentation, before diving into the core of SAM-MaGuP , which enhances SAM’s performance in polyp segmentation through the innovative MaGuP Module . This module brings unprecedented precision to weak-boundary scenarios, ensuring more effective segmentation. 2.1 SAM Overview and Limitations for Polyp Segmentation SAM comprises three key components, Image Encoder : Built on the ViT-H/16 architecture and pre-trained with Masked Autoencoders (MAE), this encoder uses windowed attention and global attention blocks to output down-sampled image embeddings. Prompt Encoder : This module processes both sparse and dense prompts, tailored with positional encoding and learned embeddings. Mask Decoder : A modified transformer decoder with dynamic mask prediction, refining the interaction between image and prompt embeddings to output prediction. While SAM excels in general object segmentation, its performance falters in weak-boundary polyp segmentation due to two major limitations: Inferior Transfer Learning : SAM’s traditional full fine-tuning approach results in overfitting or feature degradation when trained on small or limited datasets. Weak-Boundary Challenge : SAM is ill-equipped to handle ambiguous, low-contrast polyp regions that blend with the background, making it difficult to generate accurate