Evaluating the impact of preprocessing on CNN architectures: comparative analysis of a novel proposed model EyeDiagNet and existing models for eye disease detection

Abstract Purpose This study introduces a new Convolutional Neural Network (CNN) architecture, EyeDiagNet, for the classification of fundus images into four classes: normal, cataract, diabetic retinopathy (DR), and glaucoma. The performance of EyeDiagNet is compared against well-known CNN architectures to assess its effectiveness in automated eye disease classification. Methods Seven CNN models, including AlexNet, VGG, GoogleNet, ResNet, RelayNet, MobileNet, and the proposed EyeDiagNet, were trained on fundus image datasets. Both original and pre-processed datasets were utilized to analyse the impact of preprocessing on model performance. The models were evaluated using accuracy, precision, recall, and confusion matrix analysis. Results EyeDiagNet demonstrated the highest classification performance on the pre-processed dataset, achieving an accuracy of 89.97%, precision of 87.68%, and recall of 87.91%. Compared to other CNN models, EyeDiagNet exhibited superior performance, highlighting its effectiveness in distinguishing between normal, cataracts, DR, and glaucoma. The findings indicate that preprocessing enhances classification accuracy, reinforcing the importance of image enhancement techniques in deep learning-based ophthalmic disease detection. To ensure real-world applicability, the model was tested on 176 fundus images obtained from AIMS (Amrita Institute of Medical Sciences) and 144 fundus images (covering all balanced four classes: normal, cataract, diabetic retinopathy, and glaucoma) were collected from Sabitha Eye Care Hospital, Pathanamthitta, Kerala, India. In total, 320 clinically validated fundus images across four balanced classes were used for external evaluation. All data were clinically validated, and the ground truth labels were confirmed by qualified ophthalmologists and doctors, ensuring reliability and clinical accuracy of the evaluation. A key strength of this study is the inclusion of statistical significance testing to validate the performance improvements of EyeDiagNet. While many prior works report only raw accuracy values, we confirmed that the higher performance of EyeDiagNet was not due to chance. McNemar’s test demonstrated that EyeDiagNet’s accuracy gains were statistically significant over all baselines (p < 0.001 vs. MobileNet, GoogleNet, ResNet, and RelayNet; p = 0.021 vs. VGG; p = 0.042 vs. AlexNet). The inclusion of 95% confidence intervals for accuracy, precision, and recall further supports the reliability of these performance estimates. Together, these results validate the superiority of EyeDiagNet and demonstrate that its improvements are not only higher in magnitude but also statistically significant compared to baseline models. Conclusions The study demonstrates that EyeDiagNet, particularly with preprocessing, enhances automated eye disease classification. Its application has the potential to improve early detection and reduce vision impairment, especially in resource-limited settings with restricted access to specialized eye care. Translational relevance EyeDiagNet bridges the gap between deep learning research and clinical ophthalmology by providing an accurate and efficient tool for automated eye disease screening, facilitating early diagnosis and timely intervention.