Scholarlink

Accurate segmentation of colorectal polyps is crucial for early cancer detection but remains challenging due to significant variability in appearance and indistinct boundaries. We propose BiEncoder-ResMambaUNet, a dual-encoder architecture designed to enhance segmentation accuracy and generalizability across diverse clinical settings. The first encoder leverages a pre-trained EfficientNetB4 to extract high-level semantic features, refined via Depthwise Separable Convolutions and a transformer block for improved global context. The second encoder, ReSEVM, integrates Residual, Squeeze-and-Excitation (SE), and VSS Mamba blocks to capture fine-grained textures and local details. A transformer-based fusion module merges complementary encoder features, while a gated decoder with adaptive skip connections reconstructs precise polyp boundaries. Experiments on Kvasir-SEG, CVC-ClinicDB, and BKAI-IGH show consistent outperformance over existing methods, with gains of 0.3% in Dice and 0.7% in Precision, underscoring the model’s robustness and clinical reliability. github.com/Sanjana190/BiEncoder-ResMambaUnet .

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

BiEncoder-ResMambaUNet: Dual Encoder Framework Leveraging Residual Mamba Blocks and Multi-level Semantic Convolutional Features for Polyp Segmentation

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Sanjana Jhansi Ganji,
  • Panigrahi Srikanth,
  • Kaushal Sambanna,
  • Routhu Srinivasa Rao

摘要

Accurate segmentation of colorectal polyps is crucial for early cancer detection but remains challenging due to significant variability in appearance and indistinct boundaries. We propose BiEncoder-ResMambaUNet, a dual-encoder architecture designed to enhance segmentation accuracy and generalizability across diverse clinical settings. The first encoder leverages a pre-trained EfficientNetB4 to extract high-level semantic features, refined via Depthwise Separable Convolutions and a transformer block for improved global context. The second encoder, ReSEVM, integrates Residual, Squeeze-and-Excitation (SE), and VSS Mamba blocks to capture fine-grained textures and local details. A transformer-based fusion module merges complementary encoder features, while a gated decoder with adaptive skip connections reconstructs precise polyp boundaries. Experiments on Kvasir-SEG, CVC-ClinicDB, and BKAI-IGH show consistent outperformance over existing methods, with gains of 0.3% in Dice and 0.7% in Precision, underscoring the model’s robustness and clinical reliability. github.com/Sanjana190/BiEncoder-ResMambaUnet .