基于双分支加权卷积神经网络的视网膜图像质量评价方法

Retinal image quality assessment method based on dual-branch weighted convolutional neural network

近年来,眼底图像分析已成为一种直观且高效的辅助诊断技术。专家根据眼底相机捕获的视网膜图像对眼底疾病患者进行诊断,因此,眼底视网膜图像的质量对于医生提供及时且准确的疾病诊断至关重要。本文提出一种端到端的眼底视网膜图像质量评价方法,通过空间横向和纵向卷积的双分支模块进行特征提取,并对双分支所提取的特征进行加权融合,以提高模型的特征提取能力。通过自有数据集的训练,本文提出的模型准确率达到85.14%,AUC为0.9173,F1分数为0.7838。为验证模型的有效性,使用DRIMDB公开数据集进行测试,准确率达到92.11%,AUC为0.9911,F1为0.8966。实验结果表明,提出的方法对于眼底图像质量评价是有效的,具有优越的性能和高效的收敛速率。

Retinal image analysis has become an intuitive and efficient aided diagnostic technique.Experts use retinal images captured by fundus cameras to make the diagnose for patients.Consequently,high quality image plays the crucial role for doctors to provide timely and accurate disease diagnosis.An end-to-end method is proposed for evaluating fundus retinal image quality,which is trained by two-branch modules with horizontally and vertically spacial convolution,respectively.Furthermore,the extracted features of each branch carry out the weighted sum to enhance the ability of feature extraction.After training on the local dataset,the best network accuracy is 85.14%,AUC is 0.9173,and F1-score is 0.7838.Additionally,the model generalization is tested on the public DRIMDB dataset.Its accuracy,AUC,and F1-score reach 92.11%,0.9911,and 0.8966,respectively.The experimental results prove that the proposed method is effective for retinal image quality assessment with excellent performance and efficient convergence rate.