Abstract

Retinal detachment (RD) is a critical ocular condition requiring prompt diagnosis to prevent permanent vision loss. Delays in detection can lead to irreversible damage, highlighting the need for accurate and efficient diagnostic methods. This study presents a novel deep-learning framework integrated with advanced image processing techniques for RD diagnosis, utilizing optical coherence tomography (OCT) scans as the primary imaging modality. The proposed hybrid convolutional neural network (CNN) architecture combines feature fusion from multiple layers with adaptive image enhancement tailored to RD-specific characteristics. Key contributions include implementing wavelet-based noise reduction, adaptive histogram equalization, and edge-aware segmentation to preprocess OCT images effectively. An ensemble-based CNN model was also designed to extract multi-scale features and prioritize RD-relevant regions through attention mechanisms. The model achieved a classification accuracy of 97.8%, sensitivity of 97.4%, specificity of 98.6%, and an AUC-ROC of 98.7%, surpassing benchmarks in retinal diagnostics. The study also compares the proposed framework with state-of-the-art methods, demonstrating its superior performance in robustness and interpretability. These findings pave the way for deploying advanced diagnostic tools in clinical practice, enhancing early detection and treatment of retinal detachment.