Genetic Optimized Stacked Auto Encoder Based Diabetic Retinopathy Classification

Diabetic Retinopathy (DR) is a vision threatening consequence of increased glucose level in the blood. Microaneurysms and Hemorrhages are the foremost progressive symptoms of DR termed as red lesions which have greater potential of causing vision impairment. Timely investigation of digital fundus photographs collected from fundus camera helps in earlier detection of DR. Various techniques related to conventional methods have been used by researchers in the past time which involves manual feature extraction-based learning. Existing deep learning techniques like Convolutional Neural Network, customized neural networks incur cost due to huge computation and complex architecture. In this work, autoencoder based framework for DR classification is designed to achieve optimal network structure with minimum computation thereby saving the training time. The proposed model namely Genetic algorithm based Stacked Sparse Auto Encoder (GA-SSAE) involves feature extraction using two layers namely Init and Elite layer integrated with genetic algorithm (GA) and softmax classifier. The layers are trained and fine-tuned in a supervised manner using Truncated Newton Constrained optimization (TNC) method to fetch optimal weights. GA-SSAE model has been tested on the standard datasets namely Messidor, ROC and also on the real-world images. Experimental results show that the GA-SSAE model well suits the ROC and Lotus dataset with an accuracy of 98% and 95% respectively.