<p>Diabetic Retinopathy (DR) stands as the most common eye ailment among people with diabetes. It results from the weakening of blood vessels within the light-sensitive tissue situated at the back of the eye. DR leads to visual impairment and even complete loss of vision. However, vision deterioration can be prevented through early stage detection and appropriate treatment. The automated detection of non-proliferative diabetic retinopathy (NPDR), an initial stage of DR holds substantial importance for early screening, clinical diagnosis, and monitoring the progression of patients’ conditions. The proposed work introduces the novel network called Retinal Network (Ret-Net) to identify NPDR using fundus images, taken from the different datasets like Messidor, Kaggle, STARE and IDRiD. The Ret-Net introduces modified hybrid features (Mod-hyfeat) block for sustaining prominent features to identify initial stage of DR. The effectiveness of the Ret-Net is estimated using the metrics such as training, testing accuracy and its losses. Mod-hyfeat blocks in Ret-Net offers significant accuracy and loss of 97.98% and 0.01 respectively for Kaggle image sets. The Ret-Net offers a substantial enhancement in performance than the existing work. A new network called retinal network is designed dedicatedly to classify NPDR patients from healthy people (HP). The convolutional layers with varying filter sizes extract both low- and high-level features of the fundus images. An additive layer then combines the features from these different-sized convolutional filters, such as 1 × 1, 3 × 3, 5 × 5 and 7 × 7. This output is used to enhance the neuron’s learnability in the hidden layer.</p>