Retinal Disease Classification Using Convolutional Neural Networks Algorithm

Abstract

Although there is a striking demand for classifying retinal disease from Optical Coherence Tomography (OCT) images through the CNN (Convolutional Neural Network), its consummation seems to be impoverished whenever the input medical image contains the noise. The impulse noise, meaning the salt and pepper noise, is perceived as the most generic noise found in a grayscale image. Impulse noise has been identified as a very common shortcoming in the field of image processing. This paper proposed a novel technique for retinal disease classification using CNN from OCT image in noise exposure. Apart from this, seven OCT images (CNV, DME, DRUSEN, and NORMAL) were used to test. A novel filter technique has been applied to eliminate the noise that gives superb accuracy regarding Structural Similarity Index (SSIM). The prevalent filter is used to eliminate noise, which provides impoverished consummation as compared to this new technique implemented in this paper. The proposed algorithm is in three steps: the first step is to add noise in the test image. In the second step, apply a novel filter to remove the noise, and in the third step, Resnet 50 and VGG16 network is used to classify the disease. The proposed CNN ResNet 50 model utilizes to draw out distinct features of softmax & ReLU used to classify patients from their Optical Coherence Tomography (OCT) images into CNN fully connected layer. The experimental result showed that the proposed system achieved better performance by the accuracy of 98.73%.