Abstract Background Diabetic retinopathy (DR) refers to the ocular effect of diabetes. It is one of the retinal vascular diseases that can cause loss of vision. DR leads to alterations in vascular networks, including angiogenesis and capillary regression. Objective The objective of this research is to provide an effective robust and accurate automatic methodology for the early detection of DR subjects. The methodology depends on two steps: 1) Blood vessel reconstruction, enhancement, and re-continuity using written custom programs, and 2) An Artificial Neural Network (ANN) as an automatic classifier between the diabetic without diabetic retinopathy (DR) and the Mild to Moderate Non-Proliferative Diabetic Retinopathy (NPDR) subjects. Methods This approach depends on extracting the seven features, which are the most changeable features according to the morphological retinal vascular network changes. These features are the mean of the intercapillary areas as regions of interest for the largest 10 and 20 selected regions, either including or excluding the Foveal Avascular Zone (FAZ) region, FAZ perimeter, circularity index, and vascular density. The OCTA images were obtained and approved by the Ophthalmology Center in Mansoura University-Egypt. Results One hundred images were processed, distributed as follows: 40 eyes were normal, 30 eyes were diabetic without DR, and 30 eyes were NPDR subjects. The total system accuracy reached 97%. The performance parameters of the classification system for normal versus diabetic were 97.5% for sensitivity, 96.67% for specificity, and 95.2% for precision. While, the measures for a diabetic without DR versus non-proliferative DR (mild to moderate) were 96.67% for sensitivity, 96.67% for specificity, and 96.67% for precision. The maximum misclassification error was 3.33%. Conclusion The proposed methodology is capable of accurate classification of the diabetic without DR and Non-proliferative diabetic retinopathy subjects. This methodology depends on using written custom programs and a plugin for MATLAB and Fiji based Image-J software with a supervised artificial neural network. This technique achieves high accuracy, resolution, specificity, and precision with only a short time needed for diagnosis.