Visual Field Prediction for Fundus Image with Generative AI

Honggu Kang, Seonghyeon Ko, Ju-Chan Kim, Duc Tai Le, Junghyun Bum, Jong Chul Han, Hyunseung Choo

Abstract: Accurate diagnosis of glaucoma is crucial due to its high risk of blindness, but the long examination time often undermines the reliability of results by examinee's subjective factors. We propose a method to reduce examination time by generating static perimetry results with Conventional Fundus Images (CFIs) utilizing the CFI2GM technique, which leverages multimodal data. Based on data from 3,306 glaucoma patients at Samsung Medical Center in Seoul, we conducted ophthalmic image translation utilizing the Pix2Pix model. Our method, combining cGAN, L1, and SSIM loss, achieved MSE 57.9886 and PSNR 30.6057 dB. Furthermore, we received positive feedback from ophthalmologist regarding the high practical applicability of the images generated by our method. This indicates that CFI2GM can enhance the reliability of glaucoma examination results as well as reduce testing time.