An Effective Hybrid Convolutional-Modified Extreme Learning Machine in Early Stage Diabetic Retinopathy

Abstract

Diabetic retinopathy (DR) initially derives from damage to the eye blood vessels which leads to bleeding up-to permanent blindness. The severity of DR is not easily known. Therefore. it is necessary to create a system that is able to identify the severity level of DR. In this study, the identification of DR was conducted using hybrid CNN and ELM method. Hybrid CNN-ELM is useful for obtaining the most effective model in the classification system and computational time. CNN architecture is useful for extracting fundus data in image feature recognition. Several modified ELM methods (KELM, MLELM, DELM) were used to classify the severity of DR based on the results of CNN feature extraction. The classification system was tested with two datasets, namely DRIVE and Messidor. Based on the average value, the best architecture evaluation in extracting fundus data was DenseNet compared to GoogleNet, ResNet18, ResNet50, and ResNet101. Based on the computation time, the KELM method was faster than the MLELM and DELM methods, with an average time of 43 seconds. The results on the DRIVE dataset produced good evaluation values, while Messidor obtained good results on the MLELM method. It showed that the Messidor data was able to be separated well linearly. The modified CNN-MLELM method produced more stable values in both datasets with an average accuracy of 99.21%, a sensitivity of 99.29%, and a specificity of 99.21%.