Advancing Glaucoma Detection: Synthetic Image Generation via Generative Adversarial Networks and Classification with Pretrained MobileNetV2

Irreversible vision loss, generally insidious in start and lacking apparent symptoms, is usually attributed to glaucoma. Detecting glaucoma in its early stages is crucial since it may reduce disease progression. Conventional diagnostic approaches, depending on manual assessments, are notably prone to mistakes. Hence, automated glaucoma analysis acquires vital relevance for precise and prompt detection. Moreover, medical picture databases frequently display asymmetries, creating a problem. To solve these challenges, this paper presents a unique framework employing generative adversarial networks (GANs) to generate images, thus addressing dataset imbalances. Specifically., in the context of fundus images, standard approaches like image-to-image translation are applied to build synthetic fundus images and associated vascular networks, attempting to boost overall image quality and capture finer details. Gaussian filtering is originally done to pre-process the raw dataset, eliminating unwanted noise. Subsequently, GANs are applied for dataset balancing, providing synthetic images that boost classification accuracy. Optic cup segmentation is conducted using the Enhanced Level Set Algorithm. Finally, Pretrained MobileNetV2 permits accurate classification of glaucomatous images into normal and pathological categories. Experimental results illustrate the efficacy of our suggested framework, obtaining an accuracy of 98.9%, exceeding existing techniques.