Publication

Abstract : Timely and precise identification of glaucoma is essential to avert irreversible vision loss. Automating glaucoma detection is essential whilst the disease worsens over time and manual screening methods possess limitations. It facilitates early identification and treatment in order to prevent the vision impairment. Machine learning and deep learning algorithms are efficient, nevertheless, they suffer from a lack of interpretability and require a substantial amount of data and computational resources. Transfer learning proves efficient by improving scarcity of data with enhanced performance. This study seeks to examine the operational characteristics of transfer learning approaches and employ XAI (eXplainable Artificial Intelligence) method LIME to achieve clear interpretability of the features obtained by the model for prediction, so as to address the challenges related to interpretability. This study has used a combined dataset consisting of RIMONE, PAPILA, CHAKSU, and AKSHI (a dataset created by the authors) to develop ResNet-50, ResNet50-V2, VGG-16, VGG-19, and EfficientNet-V2S models. Among the developed five models, ResNet-50 and EfficientNet-V2S provide better accuracy as demonstrated in this study with an analysis using metrics and feature visualization. Both models attain a maximum test accuracy of 91.78% and the minimum test loss of, with ResNet-50 at 0.24 and EfficientNet-V2S at 0.23. In addition, these models demonstrate superior precision, specificity, sensitivity, and F1-score, all around 92. Further, feature maps and class activation maps for ResNet-50 and EfficientNet-V2S show distinct and concentrated zones of activation, indicating accurate interpretation. Thus, in this comparative study, ResNet-50 and EfficientNet-V2S are choosen as the best models for this classification task owing to their high accuracy and reliable interpretability via visual activation patterns. Consequently, the mobile app ANSAN has been integrated with the ideal clinical...