A multimodal hybrid retinopathy classification model for diagnosis and grading of diabetic disease from fundus images

Early and accurate grading is crucial in preventing vision loss associated with diabetic retinopathy (DR), a complication of diabetes mellitus (DM). In this paper, an enhancement-based deep learning framework for multi-stage retinal disease classification from retinal fundus images has been proposed. The proposed work highlights the integration of morphological feature extraction with a modified ResNet-50 architecture to grade the retinal fundus images into five stages: normal (Stage 1), mild (Stage 2), moderate (Stage 3), severe non-proliferative (Stage 4), and proliferative stages (Stage 5). To evaluate the proposed framework, three experimental configurations have been performed: (i) preprocessing with contrast-limited adaptive histogram equalization (CLAHE), Gaussian blur filtering, and an enhanced super-resolution generative adversarial network (ESRGAN) followed by data augmentation (set 1), (ii) application of CLAHE followed by data augmentation only (set 2), and (iii) without CLAHE, Gaussian blurring, and ESRGAN preprocessing and data augmentation (set 3). To remove the issue with overfitting and data generalization, we have done data augmentation. The proposed work is conducted on three public datasets, APTOS, DRIVE, DDR, and a private local retinal dataset from P.C Chatterjee Memorial Eye Hospital. The proposed method attains accuracies of 97.03%, 82.7%, and 71.32% for Sets 1, 2, and 3, respectively, where the proposed framework used two large-scale datasets (APTOS 2019 and DDR) for training, testing, and validation. Moreover, a cross-dataset evaluation was conducted on independent DRIVE and local datasets to demonstrate robustness and generalization capability, and superior performance compared to existing methods. The results confirm that the integration of CLAHE with ESRGAN significantly enhances feature representation, learning efficiency, and overall classification accuracy.