A Comparative Study of VGG16, MobileNet, and Ensemble Methods on Fundus Image-Based Cataract Detection

Artificial Intelligence (AI) in ophthalmology has been growing, driven by the increasing volume of clinical data that can be utilized for algorithm development. Image processing on fundus disease is crucial in providing visual information for cataract detection. Therefore, the application of technology that can leverage fundus images for cataract detection is still a few. In this regard, several previous studies have explored CNN algorithms for automatic cataract screening using fundus images. However, other research groups have proposed or developed new algorithms as alternatives. This research compares two CNN models, VGG16 and MobileNet, while also incorporating additional approaches such as Ensemble Learning and Hybrid Method. VGG16 and MobileNet were chosen because they are known for their high accuracy performance and computational efficiency. Source data obtained from Kaggle as many as 1000 datasets. The results of the study indicate that the model with the Ensemble Learning approach achieved the highest average accuracy, followed by MobileNet, MobileNet + SVM, VGG16, and VGG16 + SVM, with accuracies of 0.9377; 0.9150; 0.9143; 0.9129; and 0.9104, respectively. This research is expected to aid medical professionals in diagnosing cataracts more accurately and guide engineers in selecting the best machine-learning models for fundus image-based cataract detection systems.