Automatic Detection of Eye Diseases Using Enhanced Fundus Images and Pre-trained CNN Models

This study addresses the widespread blindness issue due to late diagnosis and treatment of ophthalmic diseases by proposing an automated diagnostic framework based on retinal fundus images. The system utilizes transfer learning with pre-trained deep convolutional neural networks (CNNs) to differentiate between healthy and pathological retinal fundus images (cataract, diabetic retinopathy, glaucoma) efficiently. We propose a clear four-class baseline performance for cataracts, diabetic retinopathy (DR), glaucoma, and normal instances on top of the ODIR dataset. CLAHE preprocessing with a contrast constraint, along with morphological filtering using a light approach, was employed for preprocessing. We employed a strict 70/10/20 split for training/validation/testing purposes with fixed seed values for reproducibility Model selection was conducted exclusively by means of validated metrics, with no utilization of test data. Various state-of-the-art CNN models—ResNet50, InceptionV3, GoogleNet, and MobileNet—were fine-tuned and compared regarding feature reduction and ensemble techniques. Experimental results validate that customized models achieve remarkable accuracy rates of 99.25%, 99.1%, 95.5%, and 94.7%, respectively. Among the models investigated, ResNet50 yielded the highest performance on the test data. Performance on the training/validation dataset can be examined through multiple runs, providing mean ± standard deviation summarizations: Accuracy 99.94 ± 0.57%, Balanced Accuracy 99.69 ± 0.59%, Macro-F1 99.71 ± 0.59%, Macro ROC-AUC 0.989 ± 0.001, and Macro PR-AP 0.965 ± 0.002. Class-wise performance is captured through per-class ROC and PR curve plots with 95% bootstrap confidence bands on held-out test data, accompanied by Grad-CAM attention highlighting relevant regions, which are indicative of clinical relevance. The findings demonstrate the clinical applicability of superior fundus image-based automated screening as a dependable method for the early identification and control of numerous ophthalmic conditions.