摘要
针对现有算法对微血管分割精度低、难以区分病灶区域等问题,提出一种平衡多尺度注意力网络用于分割视网膜血管。在编码阶段引入多尺度特征提取模块,提升感受野减少血管细节特征损失;在编码和解码器间增加细节增强模块,突出目标区域提高信息敏感度;设计平衡尺度注意力模块调节细节和语义特征进行最终预测,减少伪影现象。实验结果表明,在DRIVE数据集上分割准确率为96.42%、灵敏度为83.17%、特异性为98.27%,优于现有其它算法。
Aiming at the problems of low accuracy of microvessel segmentation using existing algorithms and the difficulty in distinguishing lesion areas, a balanced multi-scale attention network was proposed to segment retinal blood vessels. A multi-scale feature extraction module was introduced in the encoding stage to improve the receptive field and reduce the loss of blood vessel details. A detail enhancement module was added between the encoder and the decoder to highlight the target area and increase the information sensitivity. A balanced scale attention module was designed to adjust the details and semantic features to make final predictions and reduce artifacts. Experimental results show that the segmentation accuracy on the DRIVE dataset is 96.42%, the sensitivity is 83.17%, and the specificity is 98.27%, which is better than that of other existing algorithms.
作者
梁礼明
余洁
陈鑫
周珑颂
冯新刚
LIANG Li-ming;YU Jie;CHEN Xin;ZHOU Long-song;FENG Xin-gang(School of Electrical Engineering and Automation,Jiangxi University of Science and Technology,Ganzhou 341000,China;Department of Mechanical and Electrical Engineering,College of Applied Science,Jiangxi University of Science and Technology,Ganzhou 341000,China)
出处
《计算机工程与设计》
北大核心
2023年第2期480-487,共8页
Computer Engineering and Design
基金
国家自然科学基金项目(51365017、61463018)
江西省自然科学基金面上基金项目(20192BAB205084)
江西省教育厅科学技术研究重点基金项目(GJJ170491)。
关键词
图像处理
血管分割
空洞卷积
多尺度特征融合
校准残差模块
细节增强模块
注意力机制
image processing
blood vessel segmentation
cavity convolution
multi-scale feature fusion
calibration residual module
detail enhancement module
attention mechanism
