CSG-Mamba: A Convolutional Scoring Gating Vision State Space Network for Endoscopic Polyp Segmentation

Yuliang Wang; Jiaqi Wu; Jiaye Song; Shuxia Ren

cs.CVAug 14, 2026

CSG-Mamba: A Convolutional Scoring Gating Vision State Space Network for Endoscopic Polyp Segmentation

Authors: Yuliang Wang, Jiaqi Wu, Jiaye Song, Shuxia Ren

Organizations: School of Advanced Interdisciplinary Studies, Tiangong University, Tianjin, China

Abstract

Accurate polyp segmentation is critical for computer-aided colonoscopy, yet endoscopic images often contain low-contrast boundaries, mucosal texture interference, specular highlights, and device-dependent appearance shifts. Vision State Space Models (SSMs) provide efficient long-range modeling with linear complexity, but existing Vision Mamba segmentation models typically convert 2D features into 1D scanning sequences, which may weaken local geometric continuity and over-smooth irregular contours. We propose CSG-Mamba, a convolutional scoring gating Vision State Space network for endoscopic polyp segmentation. Built on a VM-UNet-style asymmetric U-shaped encoder-decoder, CSG-Mamba inserts a Convolutional Scoring Gating (CSG) module at the semantically rich bottleneck. CSG generates a local spatial score map through pointwise and large-kernel depthwise convolutions and recalibrates state-space features by multiplicative gating. Experiments with three random seeds show that CSG-Mamba achieves 0.9220 Dice and 15.87 HD95 on Kvasir-SEG, and 0.7418 Dice and 0.6570 mIoU on CVC-ColonDB, outperforming the baselines on most overlap and recall metrics while maintaining competitive boundary accuracy.

Figures & tables

Figure 1: Overall architecture of CSG-Mamba. The CSG module is inserted at the bottleneck, and auxiliary supervision is used only during training.

Figure 2: Detailed structures of the proposed components: (a) CSG module; (b) VSS Block.

Figure 3: Illustration of SS2D operations: (a) scan expanding unfolds 2D features into directional 1D scanning sequences; (b) scan merging restores and fuses the scanned sequences into a 2D feature representation.

Figure 4: Qualitative segmentation results. Columns show the image, ground truth, U-Net, ResUNet, PraNet, TransUNet, VM-UNet, and CSG-Mamba.

Model	mIoU	Dice	HD95 $\downarrow$	Accuracy	Sensitivity	Params
U-Net	$0.7875\pm 0.0017$	$0.8572\pm 0.0006$	$33.30\pm 1.49$	$0.9687\pm 0.0010$	$0.8804\pm 0.0068$	31.04M
ResUNet	$0.7995\pm 0.0094$	$0.8656\pm 0.0081$	$30.67\pm 3.98$	$0.9703\pm 0.0010$	$0.8878\pm 0.0053$	63.85M
PraNet	$0.8453\pm 0.0027$	$0.9006\pm 0.0014$	$20.66\pm 1.57$	$0.9764\pm 0.0017$	$0.9169\pm 0.0117$	24.51M
TransUNet	$0.8540\pm 0.0064$	$0.9082\pm 0.0061$	$17.69\pm 0.71$	$0.9776\pm 0.0005$	$0.9184\pm 0.0094$	99.04M
VM-UNet	$0.8620\pm 0.0034$	$0.9178\pm 0.0020$	$18.25\pm 1.42$	$0.9804\pm 0.0009$	$0.9284\pm 0.0042$	27.43M
VM-UNetV2	$0.8661\pm 0.0021$	$0.9180\pm 0.0016$	$17.40\pm 1.76$	$0.9809\pm 0.0010$	$0.9278\pm 0.0095$	27.63M

Table 1: In-domain evaluation results on Kvasir-SEG (mean

\pm

std).

Model	mIoU	Dice	HD95 $\downarrow$	Accuracy	Sensitivity
U-Net	$0.5922\pm 0.0093$	$0.6762\pm 0.0111$	$73.10\pm 4.79$	$0.9539\pm 0.0010$	$0.7077\pm 0.0238$
ResUNet	$0.5760\pm 0.0278$	$0.6602\pm 0.0264$	$75.48\pm 4.84$	$0.9451\pm 0.0070$	$0.7199\pm 0.0082$
PraNet	$0.7403\pm 0.0181$	$0.8163\pm 0.0167$	$34.73\pm 3.98$	$0.9684\pm 0.0032$	$0.8449\pm 0.0139$
TransUNet	$0.7497\pm 0.0031$	$0.8260\pm 0.0017$	$30.27\pm 0.80$	$0.9697\pm 0.0008$	$0.8434\pm 0.0065$
VM-UNet	$0.7689\pm 0.0050$	$0.8437\pm 0.0044$	$\mathbf{24.72\pm 1.02}$	$0.9751\pm 0.0015$	$0.8489\pm 0.0092$
VM-UNetV2	$0.7643\pm 0.0057$	$0.8364\pm 0.0066$	$30.04\pm 2.48$	$\mathbf{0.9754\pm 0.0009}$	$0.8363\pm 0.0150$

Table 2: Zero-shot test results on the full CVC-ClinicDB dataset (mean

\pm

std).

Model	mIoU	Dice	HD95 $\downarrow$	Accuracy	Sensitivity
U-Net	$0.4883\pm 0.0281$	$0.5603\pm 0.0294$	$109.21\pm 11.37$	$0.9519\pm 0.0029$	$0.5677\pm 0.0241$
ResUNet	$0.4597\pm 0.0183$	$0.5270\pm 0.0208$	$123.58\pm 9.22$	$0.9512\pm 0.0037$	$0.5359\pm 0.0205$
PraNet	$0.6029\pm 0.0179$	$0.6811\pm 0.0215$	$69.77\pm 8.07$	$0.9632\pm 0.0020$	$0.6829\pm 0.0265$
TransUNet	$0.5878\pm 0.0171$	$0.6670\pm 0.0196$	$53.42\pm 10.22$	$0.9535\pm 0.0015$	$0.6947\pm 0.0252$
VM-UNet	$0.6399\pm 0.0068$	$0.7285\pm 0.0060$	$\mathbf{39.70\pm 0.74}$	$0.9629\pm 0.0004$	$0.7372\pm 0.0101$
VM-UNetV2	$0.6436\pm 0.0082$	$0.7213\pm 0.0100$	$56.65\pm 5.10$	$0.9662\pm 0.0004$	$0.7233\pm 0.0055$

Table 3: Zero-shot test results on CVC-ColonDB (mean

\pm

std).

Figure 5: Cross-domain HD95 comparison with error bars over three random seeds.

Figure 6: Parameter efficiency scatter plot on Kvasir-SEG.

Bottleneck variant	Kvasir HD95 $\downarrow$	Kvasir Dice	ClinicDB HD95 $\downarrow$	ClinicDB Dice	ColonDB HD95 $\downarrow$	ColonDB Dice	Params
None	$18.25\pm 1.42$	$0.9178\pm 0.0020$	$\mathbf{24.72\pm 1.02}$	$0.8437\pm 0.0044$	$\mathbf{39.70\pm 0.74}$	$0.7285\pm 0.0060$	27.43M
DW $3\times 3$ + Add	$17.41\pm 0.36$	$0.9240\pm 0.0035$	$26.45\pm 1.47$	$0.8482\pm 0.0002$	$41.40\pm 4.98$	$\mathbf{0.7441\pm 0.0019}$	32.18M
DW $3\times 3$ + Gate	$15.70\pm 1.29$	$0.9237\pm 0.0031$	$27.28\pm 2.31$	$0.8465\pm 0.0061$	$46.65\pm 7.02$	$0.7348\pm 0.0083$	32.18M
DW $7\times 7$ + Add	$\mathbf{15.59\pm 0.84}$	$\mathbf{0.9260\pm 0.0033}$	$27.15\pm 2.03$	$0.8447\pm 0.0038$	$46.00\pm 1.12$	$0.7292\pm 0.0020$	32.30M
DW $7\times 7$ + Gate (CSG)	$15.87\pm 0.48$	$0.9220\pm 0.0031$	$25.62\pm 1.28$	$\mathbf{0.8493\pm 0.0074}$	$41.95\pm 3.47$	$0.7418\pm 0.0091$	32.30M

Table 4: Ablation on the CSG mechanism at the bottleneck (mean

\pm

std).

Position	Kvasir HD95 $\downarrow$	Kvasir Dice $\uparrow$	ClinicDB HD95 $\downarrow$	ClinicDB Dice $\uparrow$	ColonDB HD95 $\downarrow$	ColonDB Dice $\uparrow$	Params
none	$18.25\pm 1.42$	$0.9178\pm 0.0020$	$\mathbf{24.72\pm 1.02}$	$0.8437\pm 0.0044$	$\mathbf{39.70\pm 0.74}$	$0.7285\pm 0.0060$	27.43M
encoder only	$17.82\pm 1.83$	$0.9155\pm 0.0051$	$26.25\pm 0.84$	$0.8398\pm 0.0023$	$42.37\pm 6.27$	$\mathbf{0.7431\pm 0.0169}$	29.11M
decoder only	$16.96\pm 1.26$	$\mathbf{0.9235\pm 0.0042}$	$29.91\pm 0.87$	$0.8389\pm 0.0009$	$58.09\pm 13.13$	$0.7147\pm 0.0256$	33.94M
bottleneck only	$\mathbf{15.87\pm 0.48}$	$0.9220\pm 0.0031$	$25.62\pm 1.28$	$\mathbf{0.8493\pm 0.0074}$	$41.95\pm 3.47$	$0.7418\pm 0.0091$	32.30M
all	$19.31\pm 2.77$	$0.9145\pm 0.0129$	$27.80\pm 1.17$	$0.8418\pm 0.0067$	$49.89\pm 6.80$	$0.7353\pm 0.0101$	40.50M

Table 5: Ablation on CSG placement (mean

\pm

std).

Polyp Segmentation

Polyp segmentation uses computer vision to delineate colorectal polyps in colonoscopy images and video for detection and clinical assessment.

See all 23 papers

Polyp-VSSD: A Pre-trained Transformer with Local-Global Joint VSSD Decoding for Polyp Segmentation
Oct 23, 2026Xiaobo Zhu, Shiwei Wu, Yu An +4Polyp Segmentation Medical Image Segmentation
CAS Key Laboratory of Molecular Imaging and Beijing Key Laboratory of Molecular Imaging, Institute of Automation, Chinese Academy of Sciences, Beijing 100190, China · School of Artificial Intelligence, University of Chinese Academy of Sciences, Beijing 100049, China · School of Computer Science and Engineering, Nanjing University of Science and Technology, Nanjing, Jiangsu 210094, China +2
ASGNet: Adaptive Spectrum Guidance Network for Automatic Polyp Segmentation
Apr 16, 2026Yanguang Sun, Hengmin Zhang, Jianjun Qian +2Medical Image Analysis Polyp Segmentation
PCA Laboratory, Key Laboratory of Intelligent Perception and Systems for High-Dimensional Information of Ministry of Education, School of Computer Science and Engineering, Nanjing University of Science and Technology, Nanjing, China · School of Information Science and Engineering, East China University of Science and Technology, Shanghai, China
PolypVision: A Three-Stage Hierarchical Deep Learning Framework for Classification and Segmentation of Colorectal Polyps
Aug 11, 2026Hamidreza Bolhasani, Hamidreza Rastad, Amir Mohammad Akbari +4Medical Image Classification Polyp Segmentation
DataBioX Research, Tehran, Iran · Iran University of Science and Technology, Tehran, Iran · Tehran University of Medical Sciences, Tehran, Iran +1

Oct 23, 2026J/K move · Enter open · S save

Polyp-VSSD: A Pre-trained Transformer with Local-Global Joint VSSD Decoding for Polyp Segmentation

Xiaobo Zhu, Shiwei Wu, Yu An, Yuechen Tao+3

CAS Key Laboratory of Molecular Imaging and Beijing Key Laboratory of Molecular Imaging, Institute of Automation, Chinese Academy of Sciences, Beijing 100190, China · School of Artificial Intelligence, University of Chinese Academy of Sciences, Beijing 100049, China · School of Computer Science and Engineering, Nanjing University of Science and Technology, Nanjing, Jiangsu 210094, China+2

Polyp Segmentation Medical Image Segmentation Semantic Segmentation

Open paper AskPDF arXiv

Early detection of colorectal polyps is essential for preventing colorectal cancer, making accurate polyp segmentation a key component in computer-aided diagnosis systems. However, existing endoscopic segmentation methods often exhibit limited representational capacity and insufficient integration of global and local contextual information. To address these limitations, we propose Polyp-VSSD, a novel local-global joint decoding framework built upon a pre-trained Transformer backbone for high-precision polyp segmentation. Polyp-VSSD leverages the strong feature expressiveness of large-scale visual pre-training and unifies global semantic modeling with local fine-grained reconstruction to enhance segmentation fidelity. Extensive experiments on multiple public endoscopic datasets demonstrate that our approach significantly outperforms state-of-the-art methods in both Dice and IoU. On the EndoScene dataset, Polyp-VSSD achieves mIoU and mDice scores of 0.903 and 0.946, respectively, highlighting its robustness and strong generalization capability. This work presents an effective and scalable solution for intelligent endoscopic image analysis.

Abstract

Polyp Segmentation

Explore similar work

Polyp-VSSD: A Pre-trained Transformer with Local-Global Joint VSSD Decoding for Polyp Segmentation

ASGNet: Adaptive Spectrum Guidance Network for Automatic Polyp Segmentation

PolypVision: A Three-Stage Hierarchical Deep Learning Framework for Classification and Segmentation of Colorectal Polyps