Diabetic retinopathy (DR) is one of the primary reason of blindness over the globe. During DR, retina deteriorates continuously with the presence of lesions. It is difficult to detect DR in early stage and manual diagnosis is resource-intensive. To full-fill the requirement, two-dimensional empirical wavelet transform (2D-EWT) and composite neural network based automated DR diagnosis system is developed. The informative part of fundus image is extracted from the given image. The 2D-EWT computes different number of boundaries in magnitude spectrum followed by different modes. Therefore, a data-driven uniform boundary identification method is suggested for DR classification purpose. The acquired decomposed modes are given to the composite neural network for classification. The developed system is assessed on three different datasets MESSIDOR-1, MESSIDOR-2, and AsiaPacific Tele Ophthalmology Society (APTOS) dataset on which proposed method achieved 75%, 83.82%, and 89.28% accuracy, respectively. Efficacy of the proposed system is superior than the compared methods in terms of accuracy, area under the curve, precision, specificity, and sensitivity.