<![CDATA[Retinal diseases are a primary cause of visual impairment, requiring accurate and efficient automated classification. This study proposes a multi-class classification approach for detecting retinal disease from Optical Coherence Tomography (OCT) images using the OCT-C8 dataset, which consists of 24,000 images across eight balanced classes. Pretrained lightweight convolutional neural networks (EfficientNet-B0, ShuffleNetV2, and RegNetY-400MF) are used as feature extractors to leverage transfer learning while reducing computational cost. Instead of end-to-end deep learning, extracted features are classified using traditional machine learning models (KNN, RF, SVM, and MLP), which are more efficient for moderate-sized datasets. To improve robustness, a weighted ensemble is applied, where classifier contributions are optimized using Particle Swarm Optimization (PSO). Experimental results show that the proposed method achieves a classification accuracy of 97.35%, outperforming conventional hard and soft voting methods, while maintaining a balance between computational efficiency and performance with potential for practical deployment.]]>
Copyrights © 2026
