Scholarlink

<p>Accurate polyp segmentation is crucial for early colon cancer diagnosis. While deep learning has significantly improved segmentation performance, many state-of-the-art models suffer from high computational costs and slow inference speeds, limiting their practicality in clinical endoscopic examinations. To address this issue, we propose DilateMobileU-Net, a lightweight yet effective hybrid model that balances segmentation accuracy with computational efficiency. Our approach integrates MobileNetv2 (MV2) and DilateMobileFormer (DMF) blocks in the encoder to enhance both local feature extraction and global contextual understanding. To further explore the trade-off between segmentation accuracy and model efficiency, we design three encoder variants with different parameter sizes, allowing flexible deployment in resource-constrained environments. The decoder utilizes residual convolutional layers to refine multi-scale feature representations for precise segmentation. Extensive experiments were conducted on two public datasets, <i>i.e.,</i> Kvasir-SEG and CVC-ClinicDB datasets. Experimental results have proven that DilateMobileU-Net achieves the best results in terms of segmentation accuracy on the Kvasir-SEG dataset, with DSC and IoU of 92.56% and 88.57%, and reaches the competitive performance on the CVC-ClinicDB, with DSC and IoU of 92.45% and 87.58%, respectively. The model balances inference speed with accuracy, achieving precise polyp segmentation with minimal model parameters.</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

DilateMobileU-Net: an efficient hybrid segmentation model for polyp diagnoses

  • 期刊论文

  • 发表日期：

    2025年8月21日

查看全文
  • Huimin Guo,
  • Yin Gu,
  • Wu Du,
  • Boyang Chen,
  • Taiwei Jiao,
  • Wei Qian,
  • He Ma

摘要

Accurate polyp segmentation is crucial for early colon cancer diagnosis. While deep learning has significantly improved segmentation performance, many state-of-the-art models suffer from high computational costs and slow inference speeds, limiting their practicality in clinical endoscopic examinations. To address this issue, we propose DilateMobileU-Net, a lightweight yet effective hybrid model that balances segmentation accuracy with computational efficiency. Our approach integrates MobileNetv2 (MV2) and DilateMobileFormer (DMF) blocks in the encoder to enhance both local feature extraction and global contextual understanding. To further explore the trade-off between segmentation accuracy and model efficiency, we design three encoder variants with different parameter sizes, allowing flexible deployment in resource-constrained environments. The decoder utilizes residual convolutional layers to refine multi-scale feature representations for precise segmentation. Extensive experiments were conducted on two public datasets, i.e., Kvasir-SEG and CVC-ClinicDB datasets. Experimental results have proven that DilateMobileU-Net achieves the best results in terms of segmentation accuracy on the Kvasir-SEG dataset, with DSC and IoU of 92.56% and 88.57%, and reaches the competitive performance on the CVC-ClinicDB, with DSC and IoU of 92.45% and 87.58%, respectively. The model balances inference speed with accuracy, achieving precise polyp segmentation with minimal model parameters.