Published Paper Details:

Fundus diseases are the main cause of vision loss.Doctors use special images called fundus images to classify these diseases. Computers can help with classifying using a technology called deep learning. While deep learning techniques like CNN do not effectively remove noise. To overcome the above limitations, we introduce a hybrid model combining CNN and R-FCN. In this model, we pass the fundus images through a Convolutional Neural Network to capture a variety of low-level (edges and textures) and high-level (shapes and patterns) features that are fundamental for identifying diseases in the fundus. It is applied to feature maps developed in CNN and incorporates a mechanism of position-sensitive score map to enhance the detection of regions of interest in the image, such as lesions or vascular abnormalities in retinal images. Further, it enhances the architecture using layers of convolutions for hierarchical features in images followed by a Region Proposal Network, which creates candidate regions for classification. These regions are then classified using a set of fully connected layers, making the system highly capable of localizing and classifying retinal conditions. This innovative approach can detect disease earlier and suggest for best treatment. This paper enhances the accuracy of up to 98% in Fundus image classification.