An Automatic Ocular Disease Detection Scheme from Enhanced Fundus Images Based on Ensembling Deep CNN Networks

Millions of people around the world suffer from various ocular diseases often leading to blindness due to delayed detection and treatment. This has led to a demand for quick automated detection process from medical images including retinal fundus images. In this paper, an automatic approach for classifying normal and diseased cases from given retinal fundus images is developed based on ensembling of some suitable deep learning architectures in a transfer learning platform. Instead of directly using the raw images, it is shown that use of an enhancement technique based on adaptive histogram equalization followed by morphological operations can offer better class separation between the normal and diseased images. Some efficient deep convolutional neural network (CNN) based architectures are implemented utilizing the pre-trained weights obtained via transfer learning. In order to achieve significant improvement in the classification performance, the predictions obtained from some selected deep CNN architectures, namely ResNet50, InceptionResNetV2, EfficientNetB0 and EfficientNetB2 are combined. Comprehensive experimentation carried out on an extensive ophthalmic database show promising performance. The wide range of disease and diverse collection conditions of the fundus images affirm the suitability of the method for practical implementation.