Automated identification of cataract severity using retinal fundus images

摘要

Cataract is the most prevalent cause of blindness worldwide, which accounts for more than 51 % of overall blindness. The early detection of cataract can salvage impaired vision leading to blindness. Most of the existing cataract classification systems are based on traditional machine learning methods with hand-engineered features. The manual extraction of retinal features is generally a time-taking process and requires professional ophthalmologists. Convolutional neural network (CNN) is a widely accepted model for automatic feature extraction, but it necessitates a larger dataset to evade overfitting problems. Contrarily, classification using SVM has great generalisation power to elucidate small-sample dataset. Therefore, we proposed a hybrid model by integrating deep learning models and SVM for 4-class cataract classification. The transfer learning-based models (AlexNet, VGGNet, ResNet) are employed for automatic feature extraction and SVM performs as a recogniser. The proposed architecture evaluated on 8030 retinal images with strong feature extraction and classification techniques has achieved 95.65% of accuracy. The results of this study have verified that the proposed method outperforms conventional methods and can provide a reference for other retinal diseases.