Scholarlink

<p>Medical image segmentation is a pivotal technology in diagnostics and treatment planning, enabling precise identification and segmentation of biological structures. This study introduces ACHE-Net, a novel U-shaped network integrating adaptive decomposed convolution (ADConv) and a dynamic high-frequency feature enhancement module (DHF). ADConv balances model performance and computational overhead by dynamically adjusting the convolution kernel size based on feature channel numbers. The DHF module enhances high-frequency components using dynamic discrete wavelet transform (DWT), adjustable according to input feature resolution. Experimental results on six datasets (GlaS, BUSI, Kvasir, CVC-ClinicDB, ISIC2018 and COVID-19) demonstrate that ACHE-Net outperforms other popular models with fewer parameters, achieving significant improvements in Dice scores and HD95 metrics. Source code is available at: <a href="https://github.com/xwg111/ACHENet.">https://github.com/xwg111/ACHENet.</a></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

Enhancing medical image segmentation with adaptive convolution and dynamic high-frequency feature enhancement

  • 期刊论文

  • 发表日期：

    2025年12月26日

查看全文
  • Wenguang Xu,
  • Qian Dong,
  • Shiren Li,
  • Kamoliddin Shukurov,
  • Shu Li,
  • Maksim Davydov,
  • Jawad Hussain,
  • Guangguang Yang

摘要

Medical image segmentation is a pivotal technology in diagnostics and treatment planning, enabling precise identification and segmentation of biological structures. This study introduces ACHE-Net, a novel U-shaped network integrating adaptive decomposed convolution (ADConv) and a dynamic high-frequency feature enhancement module (DHF). ADConv balances model performance and computational overhead by dynamically adjusting the convolution kernel size based on feature channel numbers. The DHF module enhances high-frequency components using dynamic discrete wavelet transform (DWT), adjustable according to input feature resolution. Experimental results on six datasets (GlaS, BUSI, Kvasir, CVC-ClinicDB, ISIC2018 and COVID-19) demonstrate that ACHE-Net outperforms other popular models with fewer parameters, achieving significant improvements in Dice scores and HD95 metrics. Source code is available at: https://github.com/xwg111/ACHENet.