Revolutionizing Ophthalmic Diagnostics: A Deep Learning Approach to Corneal Endothelium with VGG-16

The VGG-16 algorithm consists of $\mathbf{1 6}$ substantial layers could demonstrate efficacy in evaluating corneal endothelium layers, which leads to accurate model in predicting. Indices like AUC (Area under operating curve), C-index, accuracy diagnostic rate, error rate, specificity, sensitivity, negative and positive predictive values, positive predictive values, positive likelihood ratio were calculated substantially. VGG-16 model with ensembling networks applied on corneal datasets. VGG-16 model is identified as a optimal predicting model for accurate prediction. Fifteen condition factors in each raster within study area could be examined. When ensembling these factors with VGG-16 model could contribute significance outcome on predicting corneal endothelium diseases occurrences. The results indicated that the tested deep learning technique, enhanced by the VGG-16 algorithm, outperformed alternative methods in predicting corneal events. Statistical comparisons with globally recognized corneal risk score calculators demonstrated superior classification indicators. Specifically, the AUC, C-index, and diagnostic accuracy rate exhibited a significant increase in our deep learning models with the VGG-16 algorithm compared to existing calculators, affirming the method’s robust predictive capabilities. In a parallel analysis, the impact of epochs on accuracy was assessed. The pre-trained VGG-16 model achieved an outstanding accuracy of 98.8%, emphasizing its proficiency. In comparison, a custom 5-layer CNN model, trained independently, demonstrated commendable performance, achieving an accuracy of 80% after 20 epochs. This model processed an average of 50 images in a single epoch, contributing to its overall performance metrics. The comparison between the two models highlights the trade- off between leveraging sophisticated pretrained models and developing custom models based on specific needs and available resources. Although VGG-16 excelled in accuracy, the custom CNN demonstrated competitive proficiency, showcasing the importance of selecting models tailored to the task at hand and the characteristics of the available dataset. This nuanced understanding of model performance is crucial in making informed choices in designing and deploying deep learning models for corneal risk assessments.