Scholarlink

<p>Accurate cell image segmentation remains a longstanding challenge due to the significant morphological variability and complex spatial distributions inherent in biomedical images. While U-Net architectures have demonstrated remarkable success in medical image segmentation, they often struggle to simultaneously capture long-range dependencies and multiscale semantics and preserve cross-layer consistency. In this work, we propose HOSTransNet, a hierarchical and context-aware framework designed to address these limitations through a unified feature interaction strategy. In our proposed HOSTransNet framework, we design a hierarchical context aggregation block (HCAB) to capture long-range dependencies and multiscale contextual information, utilizing progressive receptive fields. Additionally, we incorporate a spatial-channel cross transformer block (SCTB) to establish spatially aware and channel-adaptive skip connections. To mitigate the information inconsistency between skip connections and the decoder pathway, a channel-aware and lightweight decoder is proposed to align semantic features and facilitate efficient feature reconstruction. For efficient deployment, a lightweight re-parameterization strategy is applied throughout the network backbone. Comprehensive evaluations across three benchmark datasets demonstrate state-of-the-art performance: ISBI2014 (Dice: 94.58%), Glas (92.51%), and DSB2018 (90.33%). The code is available at: <a href="https://github.com/weinuan3/HOSTransNet.">https://github.com/weinuan3/HOSTransNet.</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

HOSTransNet: a Hierarchical optimized U-Net with spatial-channel cross transformer for cell image segmentation

  • 期刊论文

  • 发表日期：

    2026年3月10日

查看全文
  • Hao Wang,
  • Zhimin Lu,
  • Fuhua Ge,
  • Jundong Yang,
  • Haiyan Li,
  • Pengfei Yu

摘要

Accurate cell image segmentation remains a longstanding challenge due to the significant morphological variability and complex spatial distributions inherent in biomedical images. While U-Net architectures have demonstrated remarkable success in medical image segmentation, they often struggle to simultaneously capture long-range dependencies and multiscale semantics and preserve cross-layer consistency. In this work, we propose HOSTransNet, a hierarchical and context-aware framework designed to address these limitations through a unified feature interaction strategy. In our proposed HOSTransNet framework, we design a hierarchical context aggregation block (HCAB) to capture long-range dependencies and multiscale contextual information, utilizing progressive receptive fields. Additionally, we incorporate a spatial-channel cross transformer block (SCTB) to establish spatially aware and channel-adaptive skip connections. To mitigate the information inconsistency between skip connections and the decoder pathway, a channel-aware and lightweight decoder is proposed to align semantic features and facilitate efficient feature reconstruction. For efficient deployment, a lightweight re-parameterization strategy is applied throughout the network backbone. Comprehensive evaluations across three benchmark datasets demonstrate state-of-the-art performance: ISBI2014 (Dice: 94.58%), Glas (92.51%), and DSB2018 (90.33%). The code is available at: https://github.com/weinuan3/HOSTransNet.