Exploratory Study on Direct Prediction of Diabetes Using Deep Residual Networks

Diabetes is threatening the health of many people in the world. People may be diagnosed with diabetes only when symptoms or complications such as diabetic retinopathy start to appear. Retinal images reflect the health of the circulatory system and they are considered as a cheap and patient-friendly source of information for diagnosis purposes. Convolutional neural networks have enhanced the performance of conventional image processing techniques significantly by neglecting inconsistent feature extraction pipelines and learning informative features automatically from data. In this work we explore the possibility of using the deep residual networks as one of the state-of-the-art convolutional networks to diagnose diabetes directly from retinal images, without using any blood glucose information. The results indicate that convolutional networks are able to capture informative differences between healthy and diabetic patients and it is possible to differentiate between these two groups using only the retinal images. The performance of the proposed method is significantly higher than human experts.