Background and objective:Computer-aided cataract diagnosis (CACD) methods play a crucial role in early detection of cataract. The existing CACD methods are suffering from performance diminution due to the presence of noise in digital fundus retinal images. The lack of robustness in CACD methods against noise is a serious concern since even the presence of small noise levels may degrade the performance of cataract detection. However, noise in fundus retinal images is unavoidable due to various processes involved in the acquisition or transmission. Hence, a robust CACD method against noisy conditions is required to diagnose the cataract accurately.

Methods:In this paper, an efficient network selection based robust CACD method under additive white Gaussian noise (AWGN) is proposed. The presented method consists a set of locally- and globally-trained independent support vector networks with features extracted at various noise levels. A suitable network is then selected based on the noise level present in the input image. The automatic feature extraction technique using pre-trained convolutional neural network (CNN) is adopted to extract features from input fundus retinal images.

Results:A good-quality fundus retinal image dataset is obtained from EyePACS dataset with the use of natural image quality evaluator (NIQE) score. The synthetic noisy fundus retinal images are then generated artificially from good-quality fundus retinal images using AWGN model for effective analysis. The analysis is carried out with existing CNN based CACD methods at different noise levels. From results it is obvious that the proposed CACD method is superior in exhibiting robust performance against AWGN than existing CNN based CACD methods.

Conclusions:From the experimental results, it is clear that the proposed method show superior performance against noise when compared with existing methods in literature. The proposed method can be useful as a starting point to continue further research on CNN based robust CACD methods.

背景与目的：计算机辅助白内障诊断（CACD）方法在白内障的早期检测中起着至关重要的作用。由于数字眼底视网膜图像中存在噪声，因此现有的CACD方法正遭受性能下降的困扰。CACD方法在抗噪声方面缺乏鲁棒性，这是一个严重的问题，因为即使是很小的噪声水平也会降低白内障检测的性能。然而，由于采集或传输中涉及的各种过程，眼底视网膜图像中的噪声是不可避免的。因此，需要一种针对嘈杂条件的鲁棒CACD方法来准确诊断白内障。

方法：本文提出了一种在加性高斯白噪声（AWGN）下有效的基于网络选择的鲁棒CACD方法。提出的方法包括一组局部和全局训练的独立支持向量网络，其特征在各种噪声水平下提取。然后根据输入图像中存在的噪声水平选择合适的网络。采用预训练卷积神经网络（CNN）的自动特征提取技术从输入眼底视网膜图像中提取特征。

结果：使用自然图像质量评估器（NIQE）得分，从EyePACS数据集获得了高质量的眼底视网膜图像数据集。然后使用AWGN模型从高质量的眼底视网膜图像人工生成合成的嘈杂的眼底视网膜图像，以进行有效分析。使用现有的基于CNN的CACD方法在不同的噪声水平下进行分析。从结果显然可以看出，与现有的基于CNN的CACD方法相比，所提出的CACD方法在针对AWGN的鲁棒性能方面更胜一筹。

结论：从实验结果来看，很明显，与文献中的现有方法相比，该方法在噪声方面具有优越的性能。所提出的方法可以作为继续对基于CNN的鲁棒CACD方法进行进一步研究的起点。