Classification of retinal lesions in fundus images using atrous convolutional neural network

The long-lasting continuation of diabetes can cause damage to the nerves in the retina and may threaten the vision of the eye. These nerve lesions are ranked into micro aneurysms, hard exudates, hemorrhages, and cotton wool spots. Early-stage detection of lesions in the retina improves the betterment of successful treatments. Automatic detection of retinal lesions makes it easier for the ophthalmologists to analyze them without spending much time on manual segmentation. Image classification can be achieved by using atrous convolution method to improve the view of the lesions depending on the various lesions of retinopathy. The present technique includes the study of fundus images, normalization of shape, and normalization of size, segmentation and image classification. Atrous convolutional neural network is proposed to extract the features and classify the retinal lesions in fundus images. We achieved a classification accuracy of 0.944 which indicates the suitability of the proposed method in retinal lesion classification.