GenVFNet: Generating Visual Field From Optical Coherence Tomography Angiography by Conditional Generative Adversarial Networks

Abstract

Glaucoma investigations play an important role in glaucoma management. This study aimed to generate a visual field pattern deviation map (GenVF) from optical coherence tomography angiography (OCTA) using conditional generative adversarial networks (cGAN), namely GenVFNet. cGAN consisted of two components: generator and discriminator. The generator consisted of encoder and decoder blocks and the discriminator was the classification learner. The training data was Grad-CAM of OCTA (GradOCTA) of glaucomatous eyes. Only eyes with a visual field (VF) compatible with optic disc photograph (DP) and/or optical coherence tomography (OCT) were used. 100 eyes (90.9%) were used for training and the remaining 10 eyes (9.1%) for testing. To conform with the five severity levels in clinical diagnosis, we quantized the generated VF (GenRawVF) from GenVFNet into five levels and named the quantized image GenVF. In the experiment, GenRawVFs were compared with actual pattern deviation images (RealRawVF) using the structural similarity index measure (SSIM), normalized root mean square error (NRMSE), Fréchet inception distance (FID), and confusion matrix. The average $\times 95$ %CI SSIM, NRMSE, and FID were $0.87\times 0.02$ , $0.10\times 0.6$ , and $7.54\times 3.47$ , respectively. GenVFNet was able to construct GenVF, especially in the superior hemifield. In addition, GenVFNet was tested with VFs of eyes whose DP and/or OCT did not correlate with the available VF. According to the expert’s opinion, 81 eyes (65.8%) of GenVF were correlated with DP and/or OCT. This study demonstrated the potential of GenVFNet to construct a VF pattern deviation from the OCTA that would be a benefit for patients who cannot perform a reliable VF.

Keywords

• Artificial intelligence
• conditional generative adversarial network
• deep learning
• glaucoma
• optical coherence tomography angiography
• visual field