Scholarlink

<p>High-precision segmentation of gastrointestinal polyps in endoscopic images is crucial for early cancer screening and precise treatment. Although the U-Net network has made significant progress in endoscopic image segmentation through its encoder-decoder structure and skip connections, its linear modeling limitations cannot meet the segmentation requirements in complex scenarios. The emerging U-KAN model improves nonlinear modeling capabilities using Kolmogorov-Arnold Networks (KAN), but still faces key technical challenges in practical applications: interference from noise in low-contrast and complex backgrounds, detail loss caused by feature scale mismatches due to skip connections, and insufficient long-term dependencies modeling. To address these issues, this paper proposes an improved MAE-UKAN segmentation model that innovatively introduces a Dual-path Multi-scale Attention (DMSA) module and a Multi-scale Feature Enhancement (MSFE) module. The DMSA module integrates a Multi-scale Pyramid (MSP) and global-local feature extraction mechanisms to reduce noise interference, focus the model on lesion regions, and enhance long-term dependencies modeling. The MSFE module employs low-frequency structural enhancement and high-frequency detail extraction strategies to effectively alleviate scale mismatch issues in skip connections while preserving polyp details. Experimental results on three datasets (ES-Gastric, CVC-ClinicDB, and Kvasir-SEG) demonstrate that the MAE-UKAN model achieves Dice scores of 86.49%, 93.29%, and 87.81%, respectively, showing significant improvements over mainstream methods such as U-KAN and U-Mamba. This study provides efficient technical support for automated detection and precise segmentation of gastrointestinal polyps, with potential for clinical application extension to other gastrointestinal diseases.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

MAE-UKAN: An endoscopic images segmentation model based on dual-path multi-scale attention and feature enhancement

  • 期刊论文

  • 发表日期：

    2026年3月19日

查看全文
  • Xuekai Yin,
  • Ting Wang,
  • Dongzhi He

摘要

High-precision segmentation of gastrointestinal polyps in endoscopic images is crucial for early cancer screening and precise treatment. Although the U-Net network has made significant progress in endoscopic image segmentation through its encoder-decoder structure and skip connections, its linear modeling limitations cannot meet the segmentation requirements in complex scenarios. The emerging U-KAN model improves nonlinear modeling capabilities using Kolmogorov-Arnold Networks (KAN), but still faces key technical challenges in practical applications: interference from noise in low-contrast and complex backgrounds, detail loss caused by feature scale mismatches due to skip connections, and insufficient long-term dependencies modeling. To address these issues, this paper proposes an improved MAE-UKAN segmentation model that innovatively introduces a Dual-path Multi-scale Attention (DMSA) module and a Multi-scale Feature Enhancement (MSFE) module. The DMSA module integrates a Multi-scale Pyramid (MSP) and global-local feature extraction mechanisms to reduce noise interference, focus the model on lesion regions, and enhance long-term dependencies modeling. The MSFE module employs low-frequency structural enhancement and high-frequency detail extraction strategies to effectively alleviate scale mismatch issues in skip connections while preserving polyp details. Experimental results on three datasets (ES-Gastric, CVC-ClinicDB, and Kvasir-SEG) demonstrate that the MAE-UKAN model achieves Dice scores of 86.49%, 93.29%, and 87.81%, respectively, showing significant improvements over mainstream methods such as U-KAN and U-Mamba. This study provides efficient technical support for automated detection and precise segmentation of gastrointestinal polyps, with potential for clinical application extension to other gastrointestinal diseases.