数据与计算发展前沿 ›› 2025, Vol. 7 ›› Issue (3): 162-173.

CSTR: 32002.14.jfdc.CN10-1649/TP.2025.03.013

doi: 10.11871/jfdc.issn.2096-742X.2025.03.013

• 技术与应用 • 上一篇    下一篇

基于指数线性残差注意力与压缩激励模块的病理图像分割网络

韩铃1,*(),徐军2,洪珊珊1   

  1. 1.南京信息工程大学，自动化学院，江苏 南京 210044
    2.南京信息工程大学，人工智能学院，江苏 南京 210044
  • 收稿日期:2024-12-06 出版日期:2025-06-20 发布日期:2025-06-25
  • 通讯作者: *韩铃（E-mail: hanling_nuist@163.com）
  • 作者简介:韩铃，南京信息工程大学自动化学院，硕士研究生，主要研究方向为医学图像处理和计算病理。
    本文承担工作为整体内容的撰写及算法实现。
    HAN Ling is a master’s student at the school of Automation, Nanjing University of Information Science and Technology. Her main research areas include medical image processing and computational pathology.
    In this paper, She is responsible for writing the overall content and collecting and processing the literature.
    E-mail：hanling_nuist@163.com

The Pathological Image Segmentation Network Based on Exponential Linear Residual Attention and Squeeze-and-Excitation Module

HAN Ling1,*(),XU Jun2,HONG Shanshan1   

  1. 1. School of Automation, Nanjing University of Information Science and Technology, Nanjing, Jiangsu 210044, China
    2. School of Artificial Intelligence, Nanjing University of Information Science and Technology, Nanjing, Jiangsu 210044, China
  • Received:2024-12-06 Online:2025-06-20 Published:2025-06-25

RichHTML

0

PDF

2

摘要：

【目的】随着病理图像分析在医学中的重要性不断提高，本研究旨在开发一种高效的鼻咽癌病理图像分割方法，以提高分割精度和鲁棒性。【方法】本研究提出一种结合注意力机制与通道优化模块的分割网络SELA-CSNet。首先，设计了指数线性残差注意力单元，通过引入ELU激活函数，增强了特征表示的非线性能力，并提高了梯度传播的稳定性。随后，压缩激励模块通过自适应调整通道权重，进一步优化了网络对重要特征的关注。【结果】实验表明，本研究提出的网络结构在Dice系数上达到89.8%，显著优于传统残差网络和单一注意力机制，同时在IoU和精准度等其他指标上也表现出更优性能。【结论】该方法在当前数据集上的各项指标均优于其他先进的算法，能够有效提取关键特征，展现出较高的应用潜力。尽管其多中心泛化能力尚待验证，但已为病理图像分析提供了重要参考。

关键词: 鼻咽癌, 肿瘤微环境, 图像分割, 指数线性残差网络, 压缩激励模块, 注意力机制

Abstract:

[Objective] With the growing importance of pathological image analysis in medicine, this study aims to develop an efficient segmentation method for nasopharyngeal carcinoma pathological images to enhance segmentation accuracy and robustness. [Methods] This study proposes a segmentation network, SELA-CSNet, which integrates attention mechanisms and channel optimization modules. The exponential linear residual attention unit, incorporating the ELU activation function, enhances the nonlinear capability of feature representation and stabilizes gradient propagation. Meanwhile, the compression excitation module optimizes the network's focus on key features by adaptively adjusting channel weights. [Results] Experiments demonstrate that the proposed network architecture achieves a Dice coefficient of 89.8%, significantly outperforming traditional residual networks and single attention mechanisms. It also exhibits superior performance in other metrics such as IoU and precision. [Conclusions] SELA-CSNet surpasses other state-of-the-art algorithms in various metrics on the current datasets, demonstrating strong application potential. Although its generalizability across multi-center datasets requires further validation, it provides a significant reference for pathological image analysis.

Key words: nasopharyngeal carcinoma, tumor microenvironment, image segmentation, exponential linear residual network, squeeze-and-excitation module, attention mechanism