Applying Artificial Intelligence and Mobile Technologies to Enable Practical Screening for Diabetic Retinopathy

Abstract

With burgeoning middle class populations and changing lifestyles, diabetes is rapidly emerging as a major health concern worldwide. A related complication, Diabetic Retinopathy (DR), affects approximately 1 of 3 people with diabetes and is the leading cause of adult blindness worldwide. Since DR often goes undiagnosed, there has been great interest in using artificial intelligence in the form of Deep Learning algorithms to automatically predict DR using retina (fundus) images. However, the practical application of these algorithms is impeded by the different levels of DR severity, limited algorithm reproducibility, and a large variability in fundus imaging devices. To address these concerns, I present the development of an image processing pipeline and a neural net algorithm that automatically tests image quality (based on brightness, color, and amount of blur), rejects poor quality images, and re-formats each image into a standard image resolution. In order to create a generalized model, I used a public Kaggle database or approximately 35,000 retina images and applied a transfer learning approach using the Inception v3 architecture, to build a convolutional neural net (CNN) model that predicts referable DR. As expected, the performance of the resulting model depended on the severity of DR, with AUCs ranging from 0.96 for severe DR to 0.74 for mild DR. I further customize the model using a smaller data set of 1156 smart-phone images from our clinical partner in India. On the Kaggle dataset, the best model achieved a sensitivity = 0.85 and specificity = 0.87. On the smaller dataset, the model attained a sensitivity = 0.82 and specificity = 0.80. Finally, as a possible alternative to fundus imaging, I explore the use of iris images to explore possible correlations with diabetic retinopathy and diabetes. Using cluster analysis and generating plots with PCA and T-SNE, we observed possible clusters in using GLCM features; however, there was insufficient data from healthy individuals to be able to draw any significant conclusions.