Deep Learning Based Automated Detection System For Diagnosing Glaucoma Disease

Abstract

Glaucoma is an optic neuropathy that can cause permanent vision loss if not detected early. The goal of the present project is to design a controlled multi-architecture solution for automated glaucoma classification of retinal fundus images. A set of 7,815 images was collected across 6 public image repositories (ACRIMA, DRISHTI-GS, G1020, LAG, ORIGA, RIM-ONE) and duplicate images were eliminated at the file level by using a process of filename-label matching. This resulted in a dataset of 5,002 normal and 2,813 glaucomatous images that were divided into 6,252 training images, 781 validation images and 782 test images using two-stage stratified sampling. These models were tested: custom Basic CNN, hybrid CNN–Transformer (based on ResNet18), MobileNetV2, ResNet18, DenseNet121, EfficientNet-B0, Swin-T and squeeze-and-excitation attention-enhanced DenseNet121. The image were cropped to a size of 224 × 224, were normalized using the ImageNet statistics and were augmented in the training process by brightness-contrast, rotation and horizontal and vertical flipping. Adam optimizer, learning rate 2 × 10⁻⁴, adaptive learning-rate reduction, checkpointing and early stopping were used for optimization. The best classification accuracy (93.22%), specificity (98.20%), and precision (96.36%) were obtained by ResNet18. DenseNet121 with attention had the best probability discrimination with AUC (0.9820; 95% CI: 0.9744–0.9882) and highest sensitivity (87.59%). Swin-T and EfficientNet-B0 also achieved good accuracy with 92.84% and 92.58% respectively. The results show that the residual learning can give highly reliable normal case exclusion and channel attention can enhance glaucoma sensitivity and ranking performance. The proposed benchmark is a repeatable framework that can aid in the identification of compact and efficient architectures for computer-aided glaucoma screening.