Diabetic Retinopathy (DR) is a visual impairment caused by long-term diabetes that damages the blood vessels in the retina. It is one of the main causes of blindness worldwide, and early detection is very important to prevent progression. Therefore, existing Convolutional Neural Network (CNN)-based models have improved lesion detection, but lack semantic interpretability. While ontology-driven systems enable structured reasoning, they depend on precise lesion-level inputs, and bridging these paradigms enhances diagnostic reliability. Existing DR models have limitations in lesion-level accuracy, vessel segmentation, and semantic interpretation. To address these limitations, this research proposes a unified framework, namely, DR-MobiCB-Onto model comprising a modular DR detection pipeline integrating MobileNetV3 and Convolutional Block Attention Module (CBAM), namely DR-MobiCB with domain-specific ontology-driven reasoning for efficient lesion detection and semantic interpretation. Preprocessing to enhance the image includes bilateral filtering, CLAHEU, and Z-score normalization. The Extended Adaptive Density-Based Spatial Clustering (EADBSC) method is used to segment the Thick Blood Vessels (TBVs) in the enhanced image, and finally, lesions were detected using DR-MobiCB with dilated convolutions. Evaluation was performed on the Messidor and IDRiR datasets, obtaining 97.4% and 96.8% accuracy and AUC scores of 0.987 and 0.981, respectively.
Ali Alkwzahy (Wed,) studied this question.