A CNN-Based Framework with Focal Loss for Glaucoma Detection on Retinal Fundus Images

Optic cup-to-disc ratio (CDR) evaluation using ophthalmoscopy is the most widely used diagnostic method for detecting Glaucoma, a leading cause for blindness. CDR is often determined manually by ophthalmologists, giving rise to variability due to differences in expertise, experience, and different interpretations of subtle image features. This paper presents a novel methodology for the efficient detection of early glaucoma by harnessing image features within a precisely defined Region of Interest (ROI) centered around the optic disc. The method first segments the data to localize the optic disc using a pre-trained U-Net architecture. The generated optic disc mask is then used to detect the optic disc region in the color fundus images, which are then used to train the modified ResNet50 architecture for glaucoma screening. The proposed model was trained on the images acquired from the Artificial Intelligence for Robust Glaucoma Screening (AIROGS) challenge, conducted during ISBI 2022. The trained model is then tested on three publicly available datasets, namely ACRIMA, Drishti−GS, and RIM-ONE. The obtained results demonstrate the efficacy of our model to generalize to unseen data distributions, as evidenced by its exemplary performance on the three diverse datasets that were completely unseen during training. We also addressed the class imbalance problem in the training dataset by integrating the focal loss function into the proposed pipeline. Moreover, the proposed model has shown comparable performance with the other state-of-the-art techniques in identifying glaucomatous eyes from normal, healthy images.