融合注意力谱非局部块的视网膜图像质量分级

Retinal image quality grading for fused attention spectrum non-local blocks

视网膜图像质量评估(RIQA)是筛查糖尿病视网膜病变的关键组成部分之一。针对视网膜图像质量差异大且质量评估模型泛化能力不足等问题,提出一种融合注意力谱非局部块的多特征算法来对RIQA进行预测分级。首先采用融合光谱非局部块的ResNet50网络对输入图像进行特征提取;其次引入高效通道注意力用于提升模型对数据的表达能力,有效捕获通道间特征信息关系;再次利用特征迭代注意力融合模块对各局部特征信息融合;最后联合焦点损失和正则损失进一步提高质量分级的效果。在Eye-Quality数据集上准确率为88.59%,精确度为87.56%,敏感度和F1值分别为86.10%和86.74%。在RIQA-RFMiD数据集上准确率和F1值分别为84.22%和67.17%,仿真实验表明,文中算法对视网膜图像质量评估任务中具有较好的泛化能力。

Retinal image quality assessment(RIQA)is one of the key components of screening for diabetic retinopathy.Aiming at the problems of large differences in retinal image quality and insufficient generalization ability of quality evaluation models,a multi-feature algorithm that combines non-local blocks of the attention spectrum is proposed to predict and rank RIQA.First,the ResNet50 network of fused spectral non-local blocks is used to extract the features of the input images;Second,efficient channel attention is introduced to improve the model′s ability to express data and effectively capture the characteristic information relationship between channels;Then,the feature iterative attention fusion module is used to fuse the local feature information.Finally,the combined focus loss and regular loss further improve the effect of quality classification.On the Eye-Quality dataset,the accuracy rate is 88.59%,the precision is 87.56%,the sensitivity and F1 value are 86.10%and 86.74%,respectively.The accuracy and F1 values on the RIQA-RFMiD dataset are 84.22%and 67.17%,respectively,and simulation experiments show that the proposed algorithm has a good generalization ability for retinal image quality assessment tasks.