Glaucoma Classification Using Improved Pretrained Model

The most common cause of blindness in the world is glaucoma. It can lead to a decline in eyesight and quality of life if not addressed within the stipulated time. In computer vision, convolutional neural networks (CNNs) have achieved a viable solution. The most standard glaucoma screening method is fundus imaging due to the compromises it makes between mobility, size, and price. In this paper, we utilized the RIM-1 DL dataset in two different versions: RIM-1 Hospital and RIM-1 Randomly. To enhance the image quality and increase accuracy, we used median filtering pre-processing techniques in this work. We also used fundus images to automatically diagnose glaucoma using six different Image-net-trained models (DenseNet121, DenseNet169, VGG19, InceptionV3, ResNet50, and MobileNetV2). The six CNN architectures mentioned above are modified using normalization, inception, and global pooling layers in the model with the pre-trained classification models. Results indicate that the DenseNet-121 is the most accurate model in this work, improving accuracy in the RIM-1 Hospital dataset from 70.68 to 80.13 and in the RIM-1 randomly generated dataset from 79.31 to 87.67 after the models were fine-tuned.