Classification of cataract fundus image based on deep learning

Abstract

Cataract is a dulling or clouding of the lens inside the eye. Which is one of the most common diseases that might cause blindness. Considering the damage impact of cataract, we propose to use computer science for automatic cataract detection, which is based on the classification of retinal image. This method focuses on the feature extraction step of retinal image. Firstly, the maximum entropy method is used to preprocess the fundus images. Next, we use deep learning network which is based on Caffe to automatically extract more distinctive features of fundus images. Last, several representative classification algorithms are used to identify automatically extracted features. Comparing to features extracted by deep learning and wavelet feature extracted from retinal vascular, SVM(support vector machines) and Softmax are usedfor cataract classification. Finally, cataract images are classified into normal, slight, medium or severe four-class. Through comparing with the results of classification, the feature extracted from deep learning which is classified by Softmax get better accuracy. The results demonstrate that our research on deep learning is effective and has practical value.