Cataract Diagnosis Using Convolutional Neural Networks Classifiers. A Preliminary Study

Cataract detection systems are meant to assist ophthalmologists in providing a more accurate diagnostic in a shorter period of time. In this paper, the functionality of four classifiers obtained with different pre-trained Convolutional Neural Networks (CNN) were tested and compared in order to determine a suitable way of achieving automated cataract classification. Thus, fundus images taken from two different datasets are to be classified in each of the two classes (normal eye and cataract) or three classes (normal eye, moderate cataract and severe cataract) by using four pre-trained CNNs (AlexNet, GoogleNet, InceptionV3 and Inception-ResNet-V2). Our paper also focuses on the challenges of obtaining a proper cataract detection system and proposes different approaches to improving its functionality.