摘要
针对视网膜图像末端微小血管对比度低、分割不精确问题,提出一种融合多层次残差与三空间注意力机制的U型网络用于视网膜眼底血管分割。该网络在编码部分为了减少图像特征的丢失,引入原始图像后添加多层次残差模块。此外,为防止网络深层产生梯度弥散与特征数据冗余问题,在残差模块中进一步加入批量归一化与Dropout功能。在解码部分,采用三空间注意力机制来赋予类原始图像、下采样图像和上采样图像特征不同的权重,以增强特征纹理和位置信息,并实现微小血管的精确分割。实验结果显示,在公开的彩色眼底图像数据集上,与现有算法相比,本文算法的准确率、特异性、灵敏度和AUC分别为0.985、0.991、0.829和0.985,与金标准分割结果进行对比得到的血管图具有重要的临床参考价值。
To addresses the issues of low contrast and inaccurate segmentation of tiny vessels in retinal images,a U-shaped network incorporating multi-level residuals and three-space attention mechanism is proposed.In encoding stage,a multilevel residual module is added after inputting original images for preserving image features,and additionally,batch normalization and Dropout are integrated into the residual module to prevent vanishing gradient and feature data redundancy within the deep network.In decoding stage,a three-space attention mechanism is adopted to assign different weights to the features from the original images,down-sampled images,and up-sampled images,thus enhancing feature texture and position information,and achieving precise segmentation of tiny blood vessels.Experimental results on a public color fundus image dataset demonstrate that the proposed algorithm achieves higher accuracy(0.985),specificity(0.991),sensitivity(0.829),and AUC(0.985)than the existing algorithms.Moreover,the vessel maps obtained by the comparison with the gold standard are of significant reference value in clinic.
作者
杭益柳
张琼
邱建林
杨雨薇
HANG Yiliu;ZHANG Qiong;QIU Jianlin;YANG Yuwei(School of Computer and Information Engineering,Nantong Institute of Technology,Nantong 226000,China;School of Information Science and Technology,Nantong University,Nantong 226000,China)
出处
《中国医学物理学杂志》
CSCD
2024年第6期724-733,共10页
Chinese Journal of Medical Physics
基金
江苏省自然科学基金(BK20231337)
江苏省高等学校自然科学研究项目(21KJD210004)
南通市科技局基础科学研究项目(JC22022108,MSZ2022161,JCZ20173)
南通理工学院中青年骨干教师项目(ZQNGGJS202237,ZQNGGJS202234)。
