Scholarlink

<p>Early screening for cervical cancer is of paramount significance in reducing both its incidence and mortality rates. This paper proposes a lightweight cervical image classification model, EPA-ShuffleNet, to address the issues of excessive parameter counts and high computational complexity in existing models when deployed on mobile devices. Building upon the ShuffleNetV2 architecture, the model incorporates cross-stage dense connections and the SiLU activation function, while introducing an attention module based on Euler phase transformation to enhance feature representation capability. Experimental results demonstrate that the proposed model achieves a parameter count of merely 1.56 and a computation complexity of 1.63G FLOPs, while attaining an accuracy of 82.94% in cervical image classification tasks representing a 11.17% improvement over the baseline model. This provides an effective solution for real-time cervical cancer screening on low-computational-power terminals. Our code is available at <a href="https://github.com/Distant319/EPA-ShuffleNet">https://github.com/Distant319/EPA-ShuffleNet</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

EPA-ShuffleNet: An Efficient Cervical Screening Network with Euler Phase Attention and Dense-connections

  • 期刊论文

  • 发表日期：

    2026年5月20日

查看全文
  • Jianxiong Chen,
  • Jiawei Liang,
  • Zhao Zhang,
  • Liping Liu,
  • Xiaohan Zhang,
  • Xiaoyue Zhang

摘要

Early screening for cervical cancer is of paramount significance in reducing both its incidence and mortality rates. This paper proposes a lightweight cervical image classification model, EPA-ShuffleNet, to address the issues of excessive parameter counts and high computational complexity in existing models when deployed on mobile devices. Building upon the ShuffleNetV2 architecture, the model incorporates cross-stage dense connections and the SiLU activation function, while introducing an attention module based on Euler phase transformation to enhance feature representation capability. Experimental results demonstrate that the proposed model achieves a parameter count of merely 1.56 and a computation complexity of 1.63G FLOPs, while attaining an accuracy of 82.94% in cervical image classification tasks representing a 11.17% improvement over the baseline model. This provides an effective solution for real-time cervical cancer screening on low-computational-power terminals. Our code is available at https://github.com/Distant319/EPA-ShuffleNet.