Abstract: Objective To investigate whether ophthalmologists with different experience levels can distinguish authentic retinal photographs from those generated by Generative Adversarial Networks (GANs), so as to explore whether GAN can be applied in clinical teaching of ophthalmology. Design Diagnostic test. Participants 70 generated virtual retinal images and 70 real retinal photographs. Methods 20 participants including ophthalmologists, ophthalmology postgraduate students, and non-ophthalmologist physicians from Beijing Tongren Hospital were included in the present study. After training a model using real retinal photographs to create a high-quality image generator. Subsequently, 70 generated virtual retinal photographs are mixed in a 1:1 ratio with real retinal images. The participants, representing various levels of experience in ophthalmology, are then tasked with identifying the authenticity of the retinal photographs. Main Outcome Measures Sensitivity, specificity, and accuracy in distinguishing real and generated retinal images. Results The average sensitivity, specificity and accuracy were 0.578 (range: 0.314~0.871), 0.471 (range: 0.014~0.729), and 0.524 (range: 0.236~0.707) among the participating physicians respectively. There is no statistically significant difference in accuracy between the senior ophthalmologists and other ophthalmologist group (P>0.05). Conclusion Virtual retinal photographs generated by GAN models exhibit detailed features similar to real retinal images. In the future, these models could be utilized for producing high-quality retinal images to assist in ophthalmic education and training for primary care retinal specialists. (Ophthalmol CHN, 2024, 33: 223-225)

Key words:  artificial intelligence, generative adversarial networks, ophthalmology, teaching