Advancing AMD Detection: Dataset Design and Deep Learning Optimization for Unconstrained Retinal Images

Abstract

Age-related macular degeneration (AMD) is one of the leading causes of vision impairment worldwide, making early and accurate detection essential for effective clinical intervention. Recent advances in deep learning have demonstrated promising results in automated retinal image analysis; however, most existing approaches rely on datasets acquired under controlled conditions, limiting their generalizability to real-world clinical environments. In this paper, we propose a novel AMD dataset designed to simulate unconstrained imaging conditions, by incorporating noise, luminance variations, and device-related artifacts commonly encountered during retinal scan acquisition. Using this dataset, we conduct a comprehensive comparative evaluation of six widely adopted deep learning architectures: VGG16, VGG19, InceptionV3, MobileNetV2, ResNet50, and DenseNet. Experimental results indicate notable performance variations across models, highlighting the impact of architectural design on robustness to image degradation. Among the evaluated approaches, VGG16 achieved the best overall performance. By further optimizing this architecture through targeted training and fine-tuning strategies, the proposed system reached an accuracy of 88% in AMD detection. These findings demonstrate the effectiveness of the optimized VGG16 model and underline the importance of realistic datasets for developing reliable deep learning-based diagnostic tools for practical clinical settings.