Segmentation-Enhanced Deep Learning for AMD Detection from OCT Images

Abstract

Vision is a vital human sense, but retinal pathologies, particularly Age-Related Macular Degeneration (AMD), pose a considerable obstacle since these diseases are permanent once established. Optical Coherence Tomography (OCT) offers a powerful solution, able to reveal microscopic retinal anatomy at high resolution almost instantaneously. Yet final diagnosis remains heavily dependent on specialists, slowing the chance for timely intervention. To address this limitation, the present investigation introduces a fully automated, deep-learning pipeline designed to accelerate and refine AMD detection. Our framework proceeds in two principal phases: OCT images are first segmented using DeepLabV3+ and U-Net architectures, followed by classification with ResNet152V2 and InceptionV3 models. Experimental results analysis confirms that the segmentation step, implemented as DeepLabV3+ with a DenseNet169 backbone, delivers the best representative quality, yielding fewer misdiagnosed cases across both subsequent classifiers. In conclusion, combining feature localization with classification improves diagnostic accuracy and supports timely, reliable AMD diagnosis.