Fovea and diabetic retinopathy: Understanding the relationship using a deep interpretable classifier

Abstract

Diabetic retinopathy affects some parts of the retina, including the fovea, as it progresses to the severe stage. However, the relationship between the fovea and diabetic retinopathy progression remains unknown. Here, a new methodology is proposed to analyze the relationship between the fovea morphology and diabetic retinopathy progression. The procedure is built in four phases. First, data preparation is performed. In the second part, a deep learning model for diabetic retinopathy classification is developed. Subsequently, a score for every pixel in retinal images that map its contribution to the classification result is generated using the Local Interpretable Model-Agnostic Explanation (LIME). Finally, the generated scores are analyzed to obtain the most contributing retinal parts to the diagnosis. Our framework is developed and evaluated on retinal images from the IDRiD dataset. The advantages of our methodology are two-fold. First, our classifier correctly diagnoses diabetic retinopathy with an average accuracy of >75%, better than the state-of-the-art algorithms. Moreover, our framework reveals that the fovea areas have moderate contributions to the classification result through extensive analyses using LIME. Hence, more distinctive features like retinal lesions are probably required to build a robust classification and grading system.

Keywords

• Deep learning
• Diabetic retinopathy
• Fovea
• Interpretable classifier
• LIME