融合双注意力机制与U-net模型的视网膜血管图像分割方法
A Retinal Vessel Image Segmentation Method Incorporating Dual Attention Mechanism and U-net Model
沈学利 1王昆蓬1
作者信息
- 1. 辽宁工程技术大学 软件学院,辽宁 阜新 125105
- 折叠
摘要
视网膜血管分割对于视网膜相关疾病的诊断与治疗有重要临床价值,然而视网膜图像具有复杂性与多样性,处理具有一定难度.为此,针对视网膜血管分割精度不足等问题提出一种融合坐标注意力机制CA、新型无参注意力机制SimAM与基于编码—解码结构的U-net模型的视网膜血管分割方法.该方法通过图片灰度化、限制对比度、直方图均衡化、标准化与gamma校正的预处理操作提高血管与周围组织的对比度;采用重叠平铺Overlap-tile策略将图像分割成子块以增加数据量;在U-net模型的下采样过程中引进坐标注意力机制CA,在上采样过程中引进新型无参注意力机制SimAM,通过训练构建网络模型对测试样本进行分割,得到视网膜血管图像分割结果.该模型在DRIVE数据集上的平均准确率为97.68%,平均灵敏度为80.62%,平均特异性为98.17%,AUC系数为0.982 4;同时,与视网膜血管图像分割对照方法相比表现出更佳的分割效果.融合双注意力机制与U-net模型的分割方法可有效提高视网膜血管分割性能.
Abstract
Retinal vascular segmentation has important clinical value for the diagnosis and treatment of retinal related diseases,but retinal im-ages are complex and diverse,making its processing challenging.Therefore,to address the issue of insufficient accuracy in retinal vessel seg-mentation,a retinal vessel segmentation method is proposed that integrates coordinate attention mechanism CA and a novel parameterless at-tention mechanism SimAM with U-net model based on encoding decoding structure.This method improves the contrast between blood vessels and surrounding tissues through preprocessing operations such as image grayscale,contrast limitation,histogram equalization,standardiza-tion,and gamma correction;Using overlapping tile strategy to segment the image into sub blocks to increase the amount of data;In the downs-ampling process of the U-net model,the coordinate attention mechanism CA is introduced,and in the upsampling process,a new parameter-less attention mechanism SimAM is introduced.By training and constructing a network model to segment the test samples,the segmentation re-sults of retinal vessel images are obtained.The average accuracy of the model on the DRIVE dataset is 97.68%,the average sensitivity is 80.62%,the average specificity is 98.17%,and the AUC coefficient is 0.982 4;Meanwhile,compared with the control method for retinal vas-cular image segmentation,it shows better segmentation performance.The fusion of dual attention mechanism and U-net model segmentation method can effectively improve the performance of retinal vessel segmentation.
关键词
血管分割/注意力机制/overlap-tile策略/U-net/DRIVE数据集Key words
vessel segmentation/attention mechanism/overlap-tile strategy/U-net/DRIVE dataset引用本文复制引用
出版年
2024
国家科技期刊平台
NSTL
万方数据