[1]李春晓,周亚同,单春艳,等.融合可变形卷积与注意力机制的糖尿病视网膜多病变分割网络[J].中国医学物理学杂志,2025,42(5):596-605.[doi:10.3969/j.issn.1005-202X.2025.05.007]
　LI Chunxiao,ZHOU Yatong,SHAN Chunyan,et al.A diabetic retinopathy multi-lesion segmentation network integrating deformable convolution and attention mechanism[J].Chinese Journal of Medical Physics,2025,42(5):596-605.[doi:10.3969/j.issn.1005-202X.2025.05.007]
点击复制

融合可变形卷积与注意力机制的糖尿病视网膜多病变分割网络()

《中国医学物理学杂志》[ISSN:1005-202X/CN:44-1351/R]

卷:
42
期数:
2025年第5期
页码:
596-605
栏目:
医学影像物理
出版日期:
2025-05-20

文章信息/Info

Title:
A diabetic retinopathy multi-lesion segmentation network integrating deformable convolution and attention mechanism
文章编号:
1005-202X（2025）05-0596-10
作者:
李春晓1; 周亚同1; 单春艳2; 肖志涛3; 卜云帆1
1.河北工业大学电子信息工程学院，天津 300401；2.天津医科大学朱宪彝纪念医院肾内科，天津 300134；3.天津工业大学生命科学学院，天津 300387
Author(s):
LI Chunxiao1;  ZHOU Yatong1;  SHAN Chunyan2;  XIAO Zhitao3;  BU Yunfan1
1. School of Electronic and Information Engineering, Hebei University of Technology, Tianjin 300401, China; 2. Department of Nephrology, Tianjin Medical University Chu Hsien-I Memorial Hospital, Tianjin 300134, China; 3. School of Life Sciences, Tiangong University, Tianjin 300387, China
关键词:
糖尿病视网膜病变; 可变形小波编码; 密集特征感知与聚合; 多尺度自适应融合
Keywords:
diabetic retinopathy;  deformable convolution Haar wavelet transform encoder;  dense feature perception and aggregation;  multi scale adaptive fusion
分类号:
R318；TP391.41
DOI:
10.3969/j.issn.1005-202X.2025.05.007
文献标志码:
A
摘要:
针对糖尿病视网膜病变结构复杂、不同病变尺度差异大等问题，提出一种融合可变形卷积和注意力机制的视网膜多病变分割网络用于糖尿病视网膜多病变自动分割。首先，使用可变形小波编码模块替换原始卷积下采样编码器，以适应病变的不规则形状变化，提取有效特征信息；然后，在瓶颈层引入密集特征感知与聚合模块，通过聚合多个感受野进行多尺度特征的提取，增强深层语义信息；最后，为充分融合解码器输出，提升对边缘信息的识别精度，引入多尺度自适应融合模块对每层解码器输出进行加权，从而获取最准确的分割特征图。在DDR-RLS数据集上进行硬性渗出物、出血点、软性渗出物分割验证，结果发现所提出的网络与原有Unet相比，IoU系数分别提升0.026 2、0.051 8、0.046 5，Dice系数分别提升0.027 1、0.058 1、0.050 4，AUPR值分别提升0.042 3、0.069 1、0.073 4。
Abstract:
In view of the complex structure of diabetic retinopathy and the large differences in the scales of different lesions, a novel network which integrates deformable convolution and attention mechanism is proposed for automatic diabetic retinopathy multi-lesion segmentation. Specifically, deformable convolution Haar wavelet transform encoder takes place of the original convolutional downsampling encoder to adapt to the irregular shape changes of lesions and extract effective feature information; a dense feature perception and aggregation module is introduced at the bottleneck layer to extract multiscale features by aggregating multiple receptive fields, thus enhancing deep semantic information; and finally, in order to fully integrate the decoder output and improve the recognition accuracy of edge information, a multi scale adaptive fusion module is used to weight the decoder output of each layer for obtaining the most accurate segmentation feature map. The validation of hard percolation, bleeding point, and soft percolation segmentations on the DDR-RLS dataset reveals that the proposed network shows increases of 0.026 2, 0.051 8 and 0.046 5 in IoU coefficient, 0.027 1, 0.058 1 and 0.050 4 in Dice coefficient, and 0.042 3, 0.069 1 and 0.073 4 in AUPR value, as compared with the original Unet.

相似文献/References:

[1]蒋杰伟,雷舒陶,耿苗苗,等.融合可解释性特征的糖尿病视网膜病变自动诊断[J].中国医学物理学杂志,2022,39(5):640.[doi:DOI:10.3969/j.issn.1005-202X.2022.05.020]
　JIANG Jiewei,LEI Shutao,GENG Miaomiao,et al.Automatic diagnosis of diabetic retinopathy based on interpretable features fusion[J].Chinese Journal of Medical Physics,2022,39(5):640.[doi:DOI:10.3969/j.issn.1005-202X.2022.05.020]
[2]申思源,罗冬梅.糖尿病视网膜病变的风险揭示与关键因素分析[J].中国医学物理学杂志,2022,39(6):783.[doi:DOI:10.3969/j.issn.1005-202X.2022.06.021]
　SHEN Siyuan,LUO Dongmei.Risk disclosure and key factors analysis of diabetic retinopathy[J].Chinese Journal of Medical Physics,2022,39(5):783.[doi:DOI:10.3969/j.issn.1005-202X.2022.06.021]
[3]杨东旭,赵红东,耿立新,等.双线性非局部特征结合中继监督网络用于视网膜血管分割[J].中国医学物理学杂志,2022,39(12):1516.[doi:DOI:10.3969/j.issn.1005-202X.2022.12.010]
　YANG Dongxu,ZHAO Hongdong,GENG Lixin,et al.Bilinear non-local features combined with intermediate supervision network for retinal vessel segmentation[J].Chinese Journal of Medical Physics,2022,39(5):1516.[doi:DOI:10.3969/j.issn.1005-202X.2022.12.010]
[4]曹佳悦,罗冬梅.基于Null Importance与GS-LGBM的糖尿病视网膜病变因素分析与风险预测[J].中国医学物理学杂志,2023,40(8):1033.[doi:DOI:10.3969/j.issn.1005-202X.2023.08.018]
　CAO Jiayue,LUO Dongmei.Risk factors analysis and prediction of diabetic retinopathy based on Null Importance and GS-LGBM[J].Chinese Journal of Medical Physics,2023,40(5):1033.[doi:DOI:10.3969/j.issn.1005-202X.2023.08.018]
[5]王志鲁,池越,周亚同,等.融合密集空洞注意力金字塔和多尺度的视网膜病变分割[J].中国医学物理学杂志,2024,41(8):1000.[doi:DOI:10.3969/j.issn.1005-202X.2024.08.013]
　WANG Zhilu,CHI Yue,ZHOU Yatong,et al.Diabetic retinopathy segmentation using dense dilated attention pyramid and multi-scale features[J].Chinese Journal of Medical Physics,2024,41(5):1000.[doi:DOI:10.3969/j.issn.1005-202X.2024.08.013]
[6]王文静,张莉钏,王欣,等.融合改进Retinex图像增强与深度学习的糖尿病视网膜分类检测方法[J].中国医学物理学杂志,2024,41(9):1086.[doi:DOI:10.3969/j.issn.1005-202X.2024.09.005]
　WANG Wenjing,ZHANG Lichuan,WANG Xin,et al.Classification and detection method for diabetic retinopathy based on the combination of improved Retinex image enhancement and deep learning[J].Chinese Journal of Medical Physics,2024,41(5):1086.[doi:DOI:10.3969/j.issn.1005-202X.2024.09.005]
[7]张颖,赵祺旸,郗群.基于深度融合网络研究糖尿病视网膜病变[J].中国医学物理学杂志,2025,42(3):347.[doi:10.3969/j.issn.1005-202X.2025.03.010]
　ZHANG Ying,ZHAO Qiyang,XI Qun.Diabetic retinopathy research based on deep converged network[J].Chinese Journal of Medical Physics,2025,42(5):347.[doi:10.3969/j.issn.1005-202X.2025.03.010]
[8]梁冰雪,王文婧,王皓祺,等.面向糖尿病视网膜病变分级的多层特征关注增强网络[J].中国医学物理学杂志,2025,42(9):1174.[doi:DOI:10.3969/j.issn.1005-202X.2025.09.008]
　LIANG Bingxue,WANG Wenjing,WANG Haoqi,et al.Multi-layer feature attention enhanced network for diabetic retinopathy staging[J].Chinese Journal of Medical Physics,2025,42(5):1174.[doi:DOI:10.3969/j.issn.1005-202X.2025.09.008]

备注/Memo

备注/Memo:
【收稿日期】2024-12-14【基金项目】京津冀基础研究合作专项（J210008, 21JCZXJC00170,H2021202008）；天津市医学重点学科（专科）建设项目（TJYXZDXK-032A）【作者简介】李 春 晓 ，硕 士 ，研 究 方 向 ：医 学 图 像 处 理 ，E-mail:1694553849@qq.com【通信作者】周亚同，教授，研究方向：信息感知与机器学习，E-mail:zyt@hebut.edu.cn
更新日期/Last Update: 2025-06-03