Use of Generative Adversarial Network to Improve Glaucoma Gradability of Fundus Images

Abstract

Purpose : Low-quality retinal images may impair a clinician’s ability to detect glaucoma and preclude use of deep learning (DL) models for glaucoma identification. The study’s purpose is to assess whether ungradable images can be improved by use of a paired Generative Adversarial Network (GAN) DL model, to yield gradable images.

Methods : 3217 fundus photographs from the Diagnostic Innovations in Glaucoma Study (DIGS) and African Descent and Glaucoma Evaluation Study (ADAGES) datasets were organized into 10,000 image pairs, of one inherently low quality source image and one high quality target image from the same eye. Image quality was calculated by a previously developed quality assessment DL model trained by ground truth of expert graders. The image pairs were used to train a Pix2Pix based GAN to perform image to image translation from low to high quality. The GAN was trained for 30 cycles (epochs) using adam optimizer with learning rate 2e-4, with L1 and adversarial binary cross entropy loss functions.

Results : Generated images from the model had a mean L1 distance of 0.275 and cross entropy of 4.034 from the target high quality images. Generated images had an average brightness increase of 18.30% and contrast increase of 34.31% from the source low quality images, indicating that the GAN tends to increase the brightness and contrast of input images. Qualitatively, the GAN learns quickly, requiring one epoch to improve its output from random noise (fig 1a) to a blurry rendition of the optic nerve head (figure 1b). After 30 epochs, the GAN learns to produce images with clear retinal vasculature and detail that resembles the target images (fig 1c). The GAN seems to increase brightness and improve contrast of the optic nerve head and vasculature. In some examples, the GAN may insert vasculature into the image (fig 2b). Also, the GAN may insert cropping into the generated image, mimicking the cropping in the ground truth images (fig 2c).

Conclusions : GAN shows promise to improve human gradability of fundoscopic images in the detection of glaucoma. Further investigation must be done to explain unexpected output and assess glaucoma gradability of generated images.

This abstract was presented at the 2022 ARVO Annual Meeting, held in Denver, CO, May 1-4, 2022, and virtually.

 

Figure 1: GAN sample outputs for a single example before training (A), after one epoch of training (B) and after 30 epochs of training (C).

 

Figure 2: GAN sample outputs from the final model trained for 30 epochs.