Hybrid Deep Learning Framework for Multiclass Diabetic Retinopathy Classification from Fundus Images

Accurate grading of diabetic retinopathy from retinal fundus images is crucial for early intervention and prevention of vision loss, but automated systems have trouble with fine interclass variation and good generalization across disease stages. Existing deep learning approaches commonly rely on either convolutional or transtransformer-based former based representations alone, which limits their ability to jointly capture fine-grained local lesions and global contextual patterns. To overcome the above drawback, a hybrid classification scheme combining convolutional feature extraction with transformer based global representation learning is proposed, aided with a structured data augmentation pipeline to enhance the contrast while suppressing the noise. The model is evaluated on a publicly available multi-class diabetic retinopathy dataset consisting of five severity levels and evaluated with the help of confusion matrices, classification, accuracy curves, loss curves, and receiver operating characteristic analysis. The proposed framework obtains an overall accuracy of 99.83 percent with always high precision, recall, and F1 score for all classes and an area under the curve of 1.00, outperforming strong baselines such as VGG19 ConvNeXt V2 and InternViT. Robustness is demonstrated by stable patterns of convergence behavior minimal gap between generalization and clinically consistent patterns of confusion concentrated between neighboring levels of severity rather than distant classes. Although the proposed framework is a strong performer, the evaluation has only been done on one public dataset and additional multi-institutional validation of this technique is needed.