Glaucoma is the major reason for permanent blindness and early recognition is critical for its management. Despite the deployment of many deep networks to identify glaucoma using fundus images, their generalization performance was limited due to the lack of labelled data, intricate computations, and unique hardware requirements. However, poor-quality of fundus images hampers the accurate analysis and diagnosis. There is a critical need to create more inclusive diagnostic procedures with increased accuracy. To overcome this challenge, this work introduces a multi-stage Glaucoma detection model for detecting the cruelty of the disease progression. Initially, the fundus images are collected from the standard Glaucoma database for the study. An Ensemble CNN-based feature fusion model is utilized to extract multiple features that are suitable for the severity of Glaucoma from the collected data. Ensemble CNN-based Feature fusion model is developed through the ensembling of VGG16, ResNet, and ViT models. Then extracted features are fused and sent to the proposed multi-classification model. The proposed Multi-classifier is designed using adaptive Residual GRU, where the hyperparameters of the classification unit are tuned using the Revised Fitness-based Dung Beetle Optimizer (RF-DBO). An Enhanced meta-heuristic algorithm is derived based on the Dung Beetle Optimizer. Finally, the Glaucoma severity classes are identified from the proposed Adaptive Residual GRU (AR-GRU). The performance of the proposed multi-stage glaucoma severity detection model is evaluated on standard metrics, for higher detection accuracy and the classification rate. The accuracy of the proposed model is 97.34% achieved that is comparatively better than the state-of-the-art methods.
Kumari et al. (Fri,) studied this question.