Optimizing Feature Selection in Retinal Disease Diagnosis Using RViT-AdaMobileNetV3 and WGOA Metaheuristcs

Diabetic retinopathy (DR) is a chronic microvascular complication of diabetes and remains one of the leading causes of preventable blindness worldwide. For DR, early detection and accurate stage classification through fundus image analysis are critical for timely intervention and effective treatment planning. In this work, we present a novel hybrid classification pipeline for DR screening that integrates transformer-based vision models with optimization-guided feature selection. The architecture leverages a dual-stream design, combining the global feature extraction capability of a refined vision transformer (RViT) with the local feature capturing the strength of an adaptive MobileNetV3 (AdaMobileNetV3). This design enhances the modelling of retinal lesions, such as microaneurysms, hemorrhages, and exudates, at a local scale while concurrently capturing global contextual information. A two-stage feature selection strategy is employed, initially removing highly correlated features, followed by selecting the most discriminative subset using the weighted goose optimization algorithm (WGOA). The proposed method is evaluated on two DR datasets featuring varying class distributions, including imbalanced and multi-label scenarios. The proposed architecture outperforms baseline methods (RViT-AdaMobileNetV3 and RViT-AdaMobileNetV3+GA), achieving macro F1-scores of 0.9800 on Dataset A and 0.9764 on Dataset B, thereby demonstrating its robustness, adaptability, and translational potential for cross-dataset DR classification tasks.