Fundus image transformation to indocyanine green angiography using generative adversarial networks

A colored fundus image is an initial method for examining eye diseases. Due to its limits of visualization, diseases such as wet age-related macular degeneration (wet-AMD) and polypoidal choroidal vasculopathy (PCV) on a fundus image are difficult to distinguish using only a fundus image. An indocyanine green angiography (ICGA) image has been introduced to handle such limitations, but the acquiring process is invasive, complicated, and painful to a patient. To avoid unnecessary pain for non-PCV patients, this work presents a framework to generate an ICGA from an input fundus image. Given pair-patches of high-resolution fundus and ICGA images, they are pre-processed and trained using generative adversarial networks (GAN) to construct an image generator model. The generated patches are combined and post-processed to obtain the generated ICGA image. To assess the generated ICGA images, Peak signal-to-noise ratio (PSNR) and Fréchet inception distance (FID) are used to measure the similarity to the original fundus and the ICGA images. The results show that the generated ICGA is similar to ICGA than the original fundus and the signs of polyp, which are crucial for the circulatory system diagnosis, are exhibited. In addition, EfficientNetB5 model is built to evaluate the classification performance of the generated images for the PCV and wet-AMD. The result shows that sensitivity and specificity of our generated ICGA is greater than the original fundus.