Deep learning had become the leading methodology for detecting diabetic retinopathy (DR) from fundus images. Given a large samples of fundus images with labelled medical condition i.e., diabetic retinopathy, an efficient convolution neural network (CNN) classifier can be trained. Progress had been made previously by researchers to developed a good automatic detection of DR using deep learning architecture like convolution neural network (CNN). However, previously proposed architecture for detecting DR (DR-Net) are mainly based on previous architecture developed for natural images. Not much attention had been given on configuring DR-Net hyper-parameter i.e., depth. This paper developed a new CNN-based DR-Net architecture from scratch to detect referable diabetic retinopathy (rDR) from fundus images. This paper also report analysis of different number of DR-Net's depth configuration. Compare to previous work on DR-Net, proposed architecture is simpler in terms of number of network layers while maintaining a considerably good performance. © 2021 IEEE.