Detection of Diabetic Retinopathy Severity by Image Classification Using a Hybrid Model and Attention Mechanism

Diabetic Retinopathy (DR) is an eye condition, caused by the complications of diabetes mellitus and this could even lead to extent of vision loss. DR is an irreversible process, early diagnosis and proper treatment could help to sustain vision from any further loss. It is diagnosed through colored fundus images, which require an ophthalmologist to identify the presence of lesions in images and detect the severity of the disease. As with the exponential growth of the population, people suffering with DR were also increasing rapidly and in this case, a manual diagnosis process by clinicians would be time-intensive, tedious, and cost-consuming. An effective automated DR detection using image analysis can prevent misdiagnosis, reduce time, cost, and effort. In this work, we present a novel deep learning architecture that comprises a state-of-the-art Convolutional Neural Networks (CNN) integrated with a self-attention module recently originated from a transformers network. A self-attention mechanism was applied to the image feature maps extracted from the two state-of-art architectures ResNet and DenseNet, such that the system learns from the features extracted from both the models and with respect to each other. The proposed method has obtained a propitious performance with an accuracy of 89.09%.