Publication

Abstract : Glaucoma is an optic neurological disease, which is one of leading causes for blindness worldwide. Individuals with glaucoma do not show much of the symptoms for years, and before noticing already will be in an advanced stage of visual field loss. Building glaucoma detection systems with AI that uses fundus images are cheaper options, give us a screening phase for early detection of glaucoma. This paper focuses on comparing various deep learning models (ResNet101, Efficient-Netb3, MobileNetV3, DenseNet201, ResNest50, InceptionV4) for one step detection of glaucoma using fundus images. We have used a custom dataset AKSHI with 1255 glaucomatous and 551 normal images, and a SMDG -19 dataset that combines various benchmarking datasets with 4767 glaucomatous and 7549 normal fundus photographs. Our models demonstrate notable discriminative power, with MobileNetV3 Large achieving a F1-score of 0.87, recall of 0.86, and precision of 0.87, while maintaining an overall accuracy of 0.88 with ROC AUC of 0.94 . On the other segment we have worked on blood vessels segmentation of fundus images using U-net model achieved Jaccard Index of 0.7764.