Local Binary CNN for Diabetic Retinopathy Classification on Fundus Images
Peter Mácsik, Jarmila Pavlovičová, Jozef Goga, Slavomír Kajan
Abstract
Diabetic retinopathy (DR), is currently one of the major causes of preventable blindness, worldwide.With an early diagnosis and proper treatment of this eye disease, we can prevent the spread of diabetic retinopathy.In this paper, we propose a new alternative of local binary convolutional neural network (LBCNN) deterministic filter generation which can approximate the performance of the standard convolutional neural network (CNN) with less learnable parameters and also with less memory use, which can be helpful in systems with low-memory or low computational capacity, like smart-phones.We compare our scheme with standard CNN and LBCNN that uses stochastic filter generation strategy on retinal fundus image datasets in case of binary classification into healthy and damaged classes.These experiments are also evaluated according to the standard criteria used in medical applications, such as, overall accuracy, specificity, sensitivity and predictive values.On the small dataset (Aptos), one of our proposed LBCNN architectures outperformed all of the other deep learning models examined.