First approach to the development of an adaptive Dense GRU model for ocular disease classification with multi-head sparse attention-based feature integration

Ocular disease has been a threat to eye health and degrades its functionality over a limited time. Artificial Intelligence (AI) has been adopted in finding ocular diseases from eye fundus images, and its accuracy in defining different eye disease variations is still under research. Deep learning techniques are capable of handling the fundus images and finding the substantial features representing different ocular diseases. Still, the complexities like limited generalization, complex feature engineering arises during classification, due to complex ocular datasets. Henceforth, a novel classification model is implemented to classify the ocular disease. Initially, the required fundus images are acquired from the corresponding datasets. The obtained images are then given for extracting the features, where the network, like Graph Convolutional Neural Network (GCNN), Visual Geometry Group-19-based Auto Encoder (VGG19-AE) and Inception Network (INet) are involved in the feature engineering. Features from each network are passed to the Multi-head Sparse Attention Network (MhSA) and are fused for better training purposes. Adaptive Dense Gated Recurrent Unit (AD-GRU) is constructed in this work for ocular classification, and it identifies different ocular classes as it is trained with the fused features. In this work, a novel algorithm named the Improved Random Variable-based Swarm Bipolar Optimization Algorithm (IRV-SBOA) is presented, which performs the parameter optimization of the proposed AD-GRU model. Moreover, some measures and statistical methods are used and estimated for determining the overall performance of the model in ocular classification.