Generation of fluorescein angiography images from color fundus photographs enhances the diagnosis of central serous chorioretinopathy

Abstract Objective: To enhance the diagnosis of central serous chorioretinopathy (CSC) by generating personalized fundus fluorescein angiography (FFA) from color fundus photographs (CFP). Subjects: We collected CFP and FFA images of 249 CSC and 110 normal eyes. Methods: After image registration using I2k, we trained Pix2pix algorithm to generate FFA from CFP on Google Colab. Three AI-generated image detectors and three junior ophthalmologists assessed the authenticity of the original and generated FFA images. The diagnostic performance of three junior ophthalmologists using CFP, CFP + generated FFA, and CFP + original FFA was evaluated. Finally, the agreement of original and generated FFA in the distribution of leakage points related to the foveal region was analyzed. Results: Totally 249 CSC and 110 normal FFA images were generated. Both AI detectors and junior ophthalmologists had limited ability to discriminate between original and generated FFA images (average Kappa = 0.063 and 0.301 respectively). The average diagnostic accuracy of three junior ophthalmologists improved from 0.707 using CFP to 0.857 using CFP + generated FFA, and 0.918 using CFP + original FFA. On generated FFA, 76.7% (191/249) of CSC, but none of the normal fundus showed leakage points. The weighted kappa of leakage points distribution is 0.527 between the original and generated FFA. Conclusion: FFA generation from CFP improve the diagnostic accuracy of CSC in junior ophthalmologists. Predicted leakage points may be used in management of CSC to prevent further vision loss.