Enhancing glaucoma diagnosis: Generative adversarial networks in synthesized imagery and classification with pretrained MobileNetV2

Abstract

The disease affects the optic nerve and represents the principle reasons of irreversible vision loss, mostly asymptomatic and uncontrolled. Consequently, early and accurate diagnosis is critical to prevent or reduce its effect, however, conventional diagnostic techniques often fail to provide concrete results. In this regard, we present a new approach built on Generative Adversarial Networks (GAN) and MobileNetV2 pretrained architecture for diagnosing glaucoma. Here, the imbalance in medical imaging that comes with small datasets is addressed using GANs to generate improved quality of the fundus images. Such images are synthesized using more advanced methods assured in enhancing picture reality and fine protégé to details, specifically on the vessel network in fundus images. Gaussian Filtering is used to clean the data from noise which improve the quality of inputs. Optic cup segmentation is done by an enhanced level set algorithm, to accurately separate glaucomatous characteristics. Lastly, the MobileNetV2 model is fine-tuned to make reliable differentiation of normal and glaucoma images. The effectiveness of the proposed presentations is described in detail using the following results: accuracy of classification – 98.9 %. This approach does not only contribute to glaucoma screening but also can also reveal the benefits of the GANs and transfer learning in medical imaging.

• •
  A GAN approach to generate high-quality fundus image datasets in an attempt to minimize dataset differences.
• •
  Implemented improved Enhanced Level Set Algorithm for Optic Cup segmentation.
• •
  Built on top of the pretrained MobileNetV2 to obtain better results of glaucoma classification.