diabetic retinopathy (DR), a dangerous side effect of diabetes mellitus, can result in blindness. Early detection is essential for preserving vision, but seeing an ophthalmologist for an examination is costly and time-consuming. In this work, we introduce a novel deep learning framework for automatically identifying DR in fundus photos. Our method captures fine-grained image structures of minor retinal lesions unique to DR by introducing a more potent deep learning architecture and attention mechanisms. A substantial dataset of 15,000 fundus photos from various clinical facilities was used to train the developed algorithm, which has been shown to have a 94.7% sensitivity and a 95.3% specificity for DR detection, proving to be more effective than current cutting-edge solutions. Additionally, our model performs well when it comes to classifying the severity of DR into five clinical grades based on international clinical standards (weighted kappa score: 0.87). The approach that has been shown has the potential to be a reliable tool for DR screening and could assist doctors in diagnosing DR, particularly in telemedicine or low-resource settings..