Enhancing CNNs through the use of hand-crafted features in automated fundus image classification

Eye diseases such as diabetic retinopathy and diabetic macular edema pose a major threat in today’s world as they affect a significant portion of the global population. Therefore, it is of utmost importance to develop robust solutions that can accurately detect these diseases, especially in their early stages. However, current methods, based on hand-crafted features devised by experts, are not sufficiently accurate. Several solutions have been proposed that use deep learning techniques to improve the performance of such systems. However, they ignore the highly valuable hand-crafted features, that could contribute to more accurate prediction, which underlines the significance of our research. In this paper, we revisit the problem of combining these hand-crafted features with the features extracted by neural networks with the objective of delivering more accurate predictions. We systematically study several state-of-the-art neural networks and methods and propose a number of ways to integrate them into our framework. We show that we arrived at the conclusion that it is possible to achieve significantly better results and outperform networks that do not consider hand-crafted features using the proposed methods