Comparative Analysis of Pre-trained Deep Learning Models for Cataract Detection in Color Fundus Images

In this comparative study, we investigate the performance of nine popular pre-trained deep learning models for the classification of cataracts and normal eyes using color fundus images. The models under consideration include VGG16, VGG19, InceptionV3, DenseNet121, DenseNet201, MobileNet, MobileNetV2, ResNet50, and ResNet152. Through rigorous experimentation and analysis, we evaluate the models’ training, validation, and testing accuracies to gauge their effectiveness in this medical image classification task. Our results reveal varying degrees of accuracy across the models. To provide a comprehensive assessment, we propose further analysis avenues, including the examination of accuracy and loss curves, confusion matrices, number of model parameters, and prediction times. These supplementary analyses offer insights into the models’ capabilities and potential deployment considerations. Additionally, we recommend exploring ensemble methods and fine-tuning strategies to optimize classification accuracy. This study contributes to the understanding of utilizing deep learning models for medical image classification tasks, specifically in the context of cataract detection using fundus images.