ENSEMBLE DEEP LEARNING MODELS WITH HARD VOTING FOR ACCURATE CATARACT CLASSIFICATION USING FUNDUS IMAGES
Keywords:
Cataract Detection, Deep Learning, CNN, Hard Voting, Ensemble Learning, Fundus Imaging, Medical AI Article HistoryAbstract
Cataracts are among the world's most prevalent eye conditions and one of the main reasons for blindness, especially in older people. Immediate and proper diagnosis is important because cataracts are very curable through surgery if diagnosed in time. However, specialists are not readily available in the majority of the regions, thus leading to delays in diagnosis. To address this challenge, the latest research has greatly focused on applying automated, AI based diagnostic systems. Our study proposes an ensemble-based classification framework based on three pre trained convolutional neural networks VGG16, Xception, and DenseNet121 to classify cases of cataract and normals from retinal fundus images. 1,038 cataract and 1,074 normal images were captured from diverse public datasets, including IDRiD, Ocular Recognition, and HRF. Data augmentation techniques were applied prior to model training to prevent overfitting. Final classification utilized majority voting across the three models, enhancing stability without compromising model bias. The ensemble technique achieved optimal results with 100% accuracy, which implies its high capability in supporting cataract screening and early intervention
