Scholarlink

Utilizing deep learning for automatic polyp identification during colonoscopy procedures has greatly improved modern medical imaging. However, there are two main issues with present systems in clinical settings. Firstly, polyps are heterogeneous; they differ in size, position, and illumination, and they frequently have no distinct borders from the surrounding tissue. Second is that existing current encoder-decoder networks have trouble identifying minute visual attributes, particularly in tiny polyps. We present Multi-Scale Attention Transformer Efficient Network (MATE-Net), a novel Transformer-based architecture that combines a specific Reverse Axial Enhancement and Scale Diverse Feature Extractor with an EfficientNet backbone to overcome these constraints. This architecture preserves computational efficiency while improving feature extraction capabilities. Dual-Squeeze and Channel Spatial Excitation (Dual-SCSE) decoder is incorporated in the model to optimize spatial attention and improve segmentation precision. Comprehensive evaluation across five established polyp datasets: Kvasir-SEG, CVC-ClinicDB, CVC-300, CVC-ColonDB, and ETIS-LaribPolypDB demonstrates MATE-Net’s superior learning capacity and robust generalization capabilities in polyp segmentation tasks.

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

Multi-scale Attention Transformer Efficient Network (MATE-Net) for Polyp Segmentation

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Ananya Bose,
  • Anasua Sarkar

摘要

Utilizing deep learning for automatic polyp identification during colonoscopy procedures has greatly improved modern medical imaging. However, there are two main issues with present systems in clinical settings. Firstly, polyps are heterogeneous; they differ in size, position, and illumination, and they frequently have no distinct borders from the surrounding tissue. Second is that existing current encoder-decoder networks have trouble identifying minute visual attributes, particularly in tiny polyps. We present Multi-Scale Attention Transformer Efficient Network (MATE-Net), a novel Transformer-based architecture that combines a specific Reverse Axial Enhancement and Scale Diverse Feature Extractor with an EfficientNet backbone to overcome these constraints. This architecture preserves computational efficiency while improving feature extraction capabilities. Dual-Squeeze and Channel Spatial Excitation (Dual-SCSE) decoder is incorporated in the model to optimize spatial attention and improve segmentation precision. Comprehensive evaluation across five established polyp datasets: Kvasir-SEG, CVC-ClinicDB, CVC-300, CVC-ColonDB, and ETIS-LaribPolypDB demonstrates MATE-Net’s superior learning capacity and robust generalization capabilities in polyp segmentation tasks.