Enhancing Clinical Decision Support: A Deep Learning Approach for Automated Diagnosis of Eye Diseases from Fundus Images

• Published in: UHD Journal of Science and Technology Vol. 9; no. 2; pp. 61 - 76
• Main Authors: Ali, Ismael Abdulkareem, Mahmood, Sozan Abdulla
• Format: Journal Article
• Language: English
• Published: 24.08.2025
• ISSN: 2521-4209; 2521-4217
• DOI: 10.21928/uhdjst.v9n2y2025.pp61-76

Background and Objective: One of the most crucial sensory organs that helps the human brain receive information about the outside world is the eye. Due to its structural features, the back surface of the eye (retina) provides valuable insights about various disorders. It is essential to protect the eyes from diseases that could lead to vision impairment. If diseases affecting the retina are not identified and treated promptly, vision loss cannot be reversed. Therefore, effective automatic detection systems are necessary, as manual diagnosis is not only time-consuming, expensive, and labor-intensive but also requires a high level of expertise. To address this issue, many deep learning (DL)-based solutions have been proposed for screening retinal conditions. This study aimed to develop an effective system for the automated classification of four major eye conditions to support clinical decision-making. Methods: In this research, various convolutional neural network (CNN) architectures were applied to the dataset, and their performance was recorded. The CNN models are the common transfer learning pre-trained models on the ImageNet dataset. Finally, we developed a hybrid DL model combining DenseNet169 and MobileNetV1 to extract deep features from fundus images and perform multiclass classification into four categories: diabetic retinopathy, cataract, glaucoma, and normal fundus. Results: This hybrid approach yielded impressive results, attaining 92.99%, 93.02%, 92.85%, 92.90%, and 98.77% for accuracy, precision, recall, F1-score, and area under the curve (AUC) on a publicly available Kaggle dataset, i.e., eye disease classification. These results indicate that the hybrid approach enhances classification accuracy compared to other individual pre-trained CNN models. Conclusion: In summary, this study evaluated a substantial number of pre-trained models and developed a framework based on the top two optimal-performing models. Given that retinal image detection and diagnosis are critical for patient eye therapy and rehabilitation, our study offers an innovative framework that can function as a diagnostic aid for eye-related diseases.