Eye Disease Detection Enhancement Using a Multi-Stage Deep Learning Approach

Abstract

Eye disease detection enhancement using a multi-stage deep learning approach introduces a lightweight and multi-stage deep learning framework for the automatic detection of eye diseases using retinal OCT and fundus images. The proposed model is designed to achieve a balance between accuracy and computational efficiency, making it ideal for real-world deployment, particularly in resource-limited clinical environments. The system begins with a robust preprocessing phase that normalizes and enhances the input images while addressing issues related to rotation, translation, and illumination variations. This step ensures that the model remains invariant to positional and intensity distortions, thereby improving its robustness and consistency across diverse imaging conditions. The architecture of the model is composed of three sequential stages that work together to deliver highly accurate diagnostic predictions. In the first stage, fine-grained spatial and textural features are extracted from retinal images, allowing the network to capture subtle structural variations that often indicate early disease progression. The second stage integrates a dual-branch structure that combines convolutional feature blocks with identity mapping branches. This design helps the network simultaneously learn both low-level detailed features and high-level semantic representations, enabling a deeper and more comprehensive understanding of retinal abnormalities. The extracted features from both branches are then fused to retain discriminative information and eliminate redundancy before being passed into the final stage. In the classification stage, the merged feature representations are processed through fully connected layers and a softmax-based output unit to generate precise disease predictions. This multi-stage feature integration improves the model’s discriminative power and enhances its overall accuracy compared to conventional deep learning approaches. Experimental evaluations conducted on three benchmark datasets OCT 2017, Dataset-101, and Retinal OCT C8 demonstrated that the proposed model consistently outperforms existing architectures in terms of accuracy, precision, and recall. The results confirm that the combination of spatial, frequency, and contextual features contributes to better generalization across different types of retinal disorders. Moreover, the lightweight nature of the network significantly reduces computational load and memory usage without compromising performance. This efficiency makes the system particularly suitable for clinical environments with limited computational infrastructure, such as portable diagnostic devices or remote screening setups. Overall, the proposed multi-stage deep learning framework represents a powerful and efficient solution for automated eye disease detection. It enhances diagnostic reliability, supports early disease identification, and provides a scalable pathway toward integrating artificial intelligence into ophthalmic care and decision-support systems.