Classification of Senile Cataract Disease Using Convolutional Neural Network Method and Explainable Artificial Intelligence

Abstract

Senile cataract is a major cause of visual impairment in the elderly that requires technology-based diagnosis to improve detection efficiency and accuracy. This study aims to classify the severity of senile cataracts in eye fundus images using a deep learning ensemble model approach consisting of CNN Custom and MobileNetV2, as well as Explainable AI methods in the form of Grad-CAM. The underlying theory is the Convolutional Neural Network architecture as the image feature extraction model, plus the transfer learning principle in MobileNetV2, as well as the visual interpretation of Grad-CAM to increase the transparency of the model. The research approach is experimental, with the data coming from the Senile Cataract dataset processed through augmentation and stratified division. A Custom CNN was built with four convolution blocks while MobileNetV2 was used as the pretrained feature extractor. Both were combined in the feature fusion stage and the prediction results were visualized with Grad-CAM. The evaluation results showed that this ensemble model achieved 95.6% accuracy, 95.4% macro F1-score, and an AUC-ROC area close to 1, and provided a clinically relevant heatmap of the lens opacity area. The contribution of this research is in combining two different CNN models with an interpretive approach that bridges the need for high accuracy and transparency in image-based medical applications, with potential applications in automated diagnosis systems and future telemedicine services.