Diabetic Retinopathy is the leading cause of blindness among working age adults. Automatic detection of this disease is necessary to provide timely treatment to prevent vision loss. However, manual detection is a tedious process and requires skilled clinicians. The automated detection is a challenging task as it requires the detection of very fine features in the images such as abnormalities in the retinal blood vessels and thepresence of tiny spots of blood. The challenge is compounded by the highly uneven distribution of classes, which is typical in medical applications. We illustrate the use of Convolutional Neural Networks for Diabetic Retinopathy Detection by proposing anovel arrangement of CNNs in a tree form. To this end, we use the Kaggle Diabetic Retinopathy Dataset and achieve a Kappa score of 0.72 and an accuracy of 79% on the 5-class classification task, on par with the state-of-art Kaggle-Gold level.