Comparative Evaluation of Pretrained CNN Models for Preliminary Eye Defect Diagnosis Using Fundus Images

Early detection of the eye defect is still very difficult, especially in areas where there are no eye specialists available. In this research, the performance of several pre-trained Convolutional Neural Network (CNN) models for initial identification of eye defects based on the analysis of retinal fundus images was evaluated. An experimental comparison framework was established using the Messidor-derived data set that was retrieved from the UCI Machine Learning Repository. Four well-established architectures, VGG-16, VGG-19, InceptionResNetV2, and Xception were fine-tuned with consistent image preprocessing, training, and evaluation settings. Architectural performance was compared by way of accuracy, confusion matrices, and an overall composite score to enable the evaluation of each architecture on a balanced basis. In terms of accuracy, InceptionResNetV2 was found to be the most accurate (0.8906), whereas Xception had a similar accuracy (0.8875) as well as the highest overall evaluation score (0.7402), which indicates that it performed consistently across all evaluation measures. Overall, the results indicated that pre-trained deep Convolutional Neural Network (CNN) architectures are capable of providing a useful tool for initial assessment of eye defects via fundus images. Additionally, based on their superior performance in this task, InceptionResNetV2 and Xception appear to have practical advantages when deployed in applied diagnostic support systems. Furthermore, the present study provided additional confirmation of the reliability of transfer learning as an effective strategy for early-stage ophthalmic screening, eliminating the requirement to develop a custom model.