Comparison of CNN and KNN Methods for Cataract Classification and Detection Based on Fundus Images

Cataract is an ophthalmic condition where the lens of the eye becomes cloudy, leading to symptoms such as blurry vision, increased sensitivity to light, nearsightedness, and blindness. The World Health Organization (WHO) states that 40% of blindness cases in the world are caused by cataracts. The rate of blindness caused by cataracts in developed countries is 5% while for countries and/or remote areas it is as high as 50%. This leads to a decrease in the productivity level of the society. Therefore, early detection plays an important role in this regard as it can help patients recognize the cataract at an early stage and take action according to the level of cataract experienced. The purpose of this study is to create an automated model to detect and classify cataracts into four conditions, normal, immature, mature, and hypermature using machine learning algorithms. In order to obtain an optimal model to detecting and classifying cataract, this study compares the performance of two machine learning algorithms, Convolutional Neural Network with the proposed layer and K-Nearest Neighbor. In this study, the dataset used was 2000 fundus images which is divided into 1600 training data images and 400 test data images taken from Cicendo Hospital, Garut, West Java, Indonesia. In the previous study, the cataract classification and detection system were carried out using a 3-Layer Convolutional Neural Network by classifying two conditions, normal and cataract with an accuracy value of 95%. Model performance is showed by confusion matrix analysis which includes accuracy, precision, recall, and F1-Score. The best performance is obtained when using 5-Layer Convolutional Neural Network with 98% system accuracy with hyperparameter 100 of epoch, 64 of batch size, and 0.001 of learning rate. Meanwhile, the system accuracy obtained by the K-NN method is 97% with Euclidean Distance, k = 3 and 0° of angle orientation. In this study it can be concluded that the classification system and cataract detection through fundus images obtain good results with the Convolutional Neural Network algortima with an accuracy value of 98%.