《计算机应用》唯一官方网站 ›› 2025, Vol. 45 ›› Issue (12): 4045-4054.DOI: 10.11772/j.issn.1001-9081.2024111669

• 多媒体计算与计算机仿真 • 上一篇    下一篇

基于双路径注意力机制和多尺度信息融合的皮肤病灶图像分割

王斯豪, 张笃振, 杨昌昌   

  1. 江苏师范大学 智慧教育学院,江苏 徐州 221116
  • 收稿日期:2024-11-26 修回日期:2025-02-15 接受日期:2025-02-21 发布日期:2025-03-04 出版日期:2025-12-10
  • 通讯作者: 张笃振
  • 作者简介:王斯豪(2002—),男,江苏盐城人,硕士研究生,CCF会员,主要研究方向:深度学习、医学影像处理
    张笃振(1967—),男,江苏徐州人,副教授,博士,主要研究方向:图像处理、模式识别、机器学习
    杨昌昌(1999—),男,江苏盐城人,硕士研究生,主要研究方向:人工智能、目标检测。
  • 基金资助:
    江苏省高等学校自然科学研究面上项目(19KJB520032)

Skin lesion image segmentation based on dual-path attention mechanism and multi-scale information fusion

Sihao WANG, Duzhen ZHANG, Changchang YANG   

  1. School of Smart Education,Jiangsu Normal University,Xuzhou Jiangsu 221116,China
  • Received:2024-11-26 Revised:2025-02-15 Accepted:2025-02-21 Online:2025-03-04 Published:2025-12-10
  • Contact: Duzhen ZHANG
  • About author:WANG Sihao, born in 2002, M. S. candidate. His research interests include deep learning, medical image processing.
    ZHANG Duzhen, born in 1967, Ph. D., professor. His research interests include image processing, pattern recognition, machine learning.
    YANG Changchang, born in 1999, M. S. candidate. His research interests include artificial intelligence, target detection.
  • Supported by:
    General Project of Natural Science Research in Jiangsu Higher Education Institutions(19KJB520032)

摘要:

针对皮肤病灶边界模糊、存在毛发干扰、病灶大小不一等问题,提出一种基于双路径注意力机制和多尺度信息融合的皮肤病灶分割网络。首先,在编码器部分设计基于深度可分离卷积的残差门控注意力模块DGConv(Depthwise Gate Convolution),用于捕捉病灶区域的局部信息;其次,在网络瓶颈处设计多尺度上下文关系提取模块(MCEM),采用水平平均池化及垂直平均池化建模上下文信息,并融合残差空洞卷积金字塔模块捕获的多尺度特征进一步增强对病灶全局信息的理解;再次,在跳跃连接处设计双路径注意力模块用于细化病灶信息,并利用多尺度特征融合增强(MSFE)模块实现多阶段信息的交融,丰富当前阶段传输的细节特征信息;最后,在解码器部分设计特征融合模块(FM),以解决同阶段接受野失配的问题,并逐步融合编码器输出和跳跃连接传递的特征信息得到最终的分割结果。该网络在ISIC2017(International Skin Imaging Collaboration)和ISIC2018数据集上的实验结果表明,相较于皮肤病灶分割方面表现次优的网络,所提网络的Dice指标分别提高了0.09和1.09个百分点,交并比(IoU)指标分别提高0.14和1.76个百分点;与经典U-Net相比,Dice指标分别提高5.13和3.84个百分点,IoU指标分别提高了7.74和6.04个百分点。充分说明所提网络的先进性与有效性。

关键词: 皮肤病灶分割, 注意力机制, 空洞卷积, 多尺度信息, 特征融合

Abstract:

To address issues of blurred skin lesion boundaries, hair interference, and varying lesion sizes, a skin lesion segmentation network based on dual-path attention mechanism and multi-scale information fusion was proposed. Firstly, a residual gated attention module based on depthwise separable convolution, named DGConv (Depthwise Gate Convolution), was designed in the encoder to capture local lesion information. Secondly, a Multi-scale Contextual relationship Extraction Module (MCEM) was introduced at the bottleneck, which employed horizontal average pooling and vertical average pooling to model contextual information, and integrated multi-scale features captured by residual dilated convolution pyramid module to further enhance global lesion information understanding. Thirdly, a dual-path attention module was designed at the skip connections to refine lesion details, and a Multi-Scale feature Fusion Enhancement (MSFE) module was utilized to enrich feature details transmitted in current stage by fusing multi-stage information. Finally, a feature Fusion Module (FM) was designed in the decoder to solve receptive field mismatch problem at the same stage, progressively combining encoder outputs and feature information transferred by skip connection to achieve the final segmentation results. Experiments on the ISIC2017 (International Skin Imaging Collaboration) and ISIC2018 datasets demonstrated that the proposed network outperformed suboptimal networks in skin lesion segmentation, with the Dice improvements of 0.09 and 1.09 percentage points, respectively, and the Intersection over Union (IoU) improvements of 0.14 and 1.76 percentage points, respectively. Compared to the classical U-Net, the Dice was improved by 5.13 and 3.84 percentage points, respectively, and the IoU was improved by 7.74 and 6.04 percentage points, respectively, fully proving the advantage and effectiveness of the proposed network.

Key words: skin lesion segmentation, attention mechanism, dilated convolution, multi-scale information, feature fusion

中图分类号: