Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

EndoCaver: Handling Fog, Blur and Glare in Endoscopic Images via Joint Deblurring-Segmentation

Zhuoyu Wu, Wenhui Ou, Pei-Sze Tan, Jiayan Yang, Wenqi Fang, Zheng Wang, Raphaël C. -W. Phan

TL;DR

EndoCaver tackles the challenge of degraded endoscopic imagery (fog, blur, glare) by proposing a lightweight transformer with a unidirectional dual-decoder that jointly performs deblurring and polyp segmentation. It introduces a Global Attention Module for cross-scale feature fusion, a Deblurring-Segmentation Aligner to share restoration cues across tasks, and a cosine-annealing loss scheduler (LoCoS) to stabilize multi-task optimization. The model achieves state-of-the-art Dice scores on clean data (0.922) and under severe degradation (0.889) on Kvasir-SEG while using only 7.81M parameters and 11.86 GMAC, enabling efficient on-device deployment. Ablation studies confirm the necessity of LoCoS, DSA, and GAM for strong cross-task feature sharing and robust performance under distribution shifts. Overall, EndoCaver offers a practical, computation-efficient solution for real-time endoscopic analysis with improved robustness to common imaging degradations.

Abstract

Endoscopic image analysis is vital for colorectal cancer screening, yet real-world conditions often suffer from lens fogging, motion blur, and specular highlights, which severely compromise automated polyp detection. We propose EndoCaver, a lightweight transformer with a unidirectional-guided dual-decoder architecture, enabling joint multi-task capability for image deblurring and segmentation while significantly reducing computational complexity and model parameters. Specifically, it integrates a Global Attention Module (GAM) for cross-scale aggregation, a Deblurring-Segmentation Aligner (DSA) to transfer restoration cues, and a cosine-based scheduler (LoCoS) for stable multi-task optimisation. Experiments on the Kvasir-SEG dataset show that EndoCaver achieves 0.922 Dice on clean data and 0.889 under severe image degradation, surpassing state-of-the-art methods while reducing model parameters by 90%. These results demonstrate its efficiency and robustness, making it well-suited for on-device clinical deployment. Code is available at https://github.com/ReaganWu/EndoCaver.

EndoCaver: Handling Fog, Blur and Glare in Endoscopic Images via Joint Deblurring-Segmentation

TL;DR

EndoCaver tackles the challenge of degraded endoscopic imagery (fog, blur, glare) by proposing a lightweight transformer with a unidirectional dual-decoder that jointly performs deblurring and polyp segmentation. It introduces a Global Attention Module for cross-scale feature fusion, a Deblurring-Segmentation Aligner to share restoration cues across tasks, and a cosine-annealing loss scheduler (LoCoS) to stabilize multi-task optimization. The model achieves state-of-the-art Dice scores on clean data (0.922) and under severe degradation (0.889) on Kvasir-SEG while using only 7.81M parameters and 11.86 GMAC, enabling efficient on-device deployment. Ablation studies confirm the necessity of LoCoS, DSA, and GAM for strong cross-task feature sharing and robust performance under distribution shifts. Overall, EndoCaver offers a practical, computation-efficient solution for real-time endoscopic analysis with improved robustness to common imaging degradations.

Abstract

Endoscopic image analysis is vital for colorectal cancer screening, yet real-world conditions often suffer from lens fogging, motion blur, and specular highlights, which severely compromise automated polyp detection. We propose EndoCaver, a lightweight transformer with a unidirectional-guided dual-decoder architecture, enabling joint multi-task capability for image deblurring and segmentation while significantly reducing computational complexity and model parameters. Specifically, it integrates a Global Attention Module (GAM) for cross-scale aggregation, a Deblurring-Segmentation Aligner (DSA) to transfer restoration cues, and a cosine-based scheduler (LoCoS) for stable multi-task optimisation. Experiments on the Kvasir-SEG dataset show that EndoCaver achieves 0.922 Dice on clean data and 0.889 under severe image degradation, surpassing state-of-the-art methods while reducing model parameters by 90%. These results demonstrate its efficiency and robustness, making it well-suited for on-device clinical deployment. Code is available at https://github.com/ReaganWu/EndoCaver.
Paper Structure (11 sections, 10 equations, 3 figures, 3 tables)

This paper contains 11 sections, 10 equations, 3 figures, 3 tables.

Table of Contents

  1. Introduction
  2. Methodology
  3. Overall Architecture
  4. Global Attention Module
  5. Deblurring-Segmentation Aligner
  6. Loss with Cosine Annealing (LoCoS)
  7. Experiments and Results
  8. Experimental Setup
  9. Datasets and Metrics
  10. Results and Discussions
  11. CONCLUSION

Figures (3)

  • Figure 1: Comparison of different model designs. (a–d) illustrate representative existing forms, while (e) shows our proposed Unidirectional-Guided Dual-Decoder, the core of our work. It reduces overall model parameters and computational cost while enabling higher segmentation accuracy.
  • Figure 2: Overall architecture of EndoCaver. (a) End-to-end framework with MiT encoder, GAM, and dual decoders. (b) Global Attention Module (GAM) for lightweight cross-scale aggregation. (c) Deblurring-Segmentation Aligner (DSA) for cross-task feature alignment. For decoder blocks, LiDeconv employs transposed convolution with BatchNorm and SiLU for recovery, followed by a reverse residual bottleneck sandler2018mobilenetv2 for feature extraction. LiViT only replaces the bottleneck with MobileViT mehta2021mobilevit.
  • Figure 3: The qualitative results. Our proposed EndoCaver is labelled in red. The following model corresponds to the existing model design from Fig.\ref{['fig:intro_pic']} (d) to (a).