Early and accurate detection of eye diseases is crucial for preventing vision loss and enabling early treatment. Retinal fundus photography plays a vital role in this, as it offers a convenient, non-invasive, and cost-effective method to examine the innermost surface of the eye. While useful as external diagnostic references, manual interpretation of fundus images is often time-consuming, subjective, and inconsistent. To address these challenges, this study proposes a deep learning-based method for classifying retinal fundus images into cases of diabetic retinopathy (DR), glaucoma, cataract, age-related macular degeneration (AMD), and normal. We compare four popular CNN models a base CNN, ResNet50, DenseNet121, and EfficientNet-B3 trained and validated on a curated dataset of labeled images using common pre-processing and augmentation techniques. Model performance is assessed using accuracy, precision, recall, F1-score, AUC, confusion matrices, and disease-specific analyses. The results show that transfer learning methods, particularly EfficientNet-B3, outperform the baseline CNN. Our findings suggest that deep learning has the potential to enhance retinal analysis across multiple diseases, assist ophthalmologists, and reduce their workload by enabling large-scale screening for early detection of sight-threatening conditions.