IDENTIFICATION OF DIABETIC DISEASES THROUGH RETINAL SCAN BY USING CONVOLUTIONAL NEURAL NETWORK

Abstract

Finding the diabetic disease using naked eye is crucial and intervention for human at early stage is important for early clinical and interventions according to medical treatment. To safely visualize these conditions, optical coherence tomography (OCT) is widely preferred as a non-invasive, non-contact imaging modality. Given the widespread shortage of specialized diagnostic technology and clinical support. To address this need, we developed a parameterized, lightweight framework that fuses a CNN and a transformer for multi-class retinal disease classification. By leveraging this hybrid design, the CNN captures fine-grained local lessions while the transformer encoder models long range dependencies across the entire OCT image, significantly boosting diagnostic sensitivity. This pipeline is further reinforced by a specialized convolutional block designed to maximize feature representation with a low parameter footprint. We evaluated our proposed framework against several bassline architectures. On the OCT-c8 dataset, our model achieved the highest accuracy score of 0.9800 alongside competitive recall, while simultaneously utilizing the fewest parameters and requiring the shortest pre-image inference latency. Furthermore, evaluation on the broader OCT2017 dataset demonstrated that our model outperforms on four stage, recent state of the art architecture and matches the performance of a fifth, achieving a remarkable average accuracy, precision, recall, specificity, and F1-score of 0.9985, 0.9970, 0.9990, and 0.9970, respectively. These performance metrics were achieved with a highly compact footprint. The model requires only 1.28 million parameters, enabling a rapid average processing speed of 2.5 milliseconds per image scan.

Keywords : Diabetic disease, Retinal illness, eyes diseases Computer vision, Deep learning, QW kappa metric, Deterioration.