| CPC G06T 7/0012 (2013.01) [G06T 5/50 (2013.01); G06T 7/11 (2017.01); G06V 10/25 (2022.01); G06V 10/44 (2022.01); G06V 10/46 (2022.01); G06V 10/761 (2022.01); G06T 2207/20084 (2013.01); G06T 2207/20221 (2013.01); G06T 2207/30104 (2013.01)] | 19 Claims |
|
1. A computer implemented method of generating at least one synthetic fundus image, comprising:
extracting and preserving at least one real fundus region from at least one real image of a real human fundus;
generating at least one synthetic fundus image comprising a synthetic human fundus and the preserved at least one real fundus region,
wherein an identity of the real human fundus is non-determinable from the at least one synthetic fundus image;
designating pairs of images, each pair including the one real image and the at least one synthetic fundus image;
feeding the at least one real image and the at least one synthetic fundus image into a machine learning model trained to recognize fundus regions to obtain an outcome of a similarity value denoting an amount of similarity between the at least one real image and the at least one synthetic fundus image, and
verifying that the at least one synthetic fundus image does not depict the real human fundus when the similarity value is below a threshold,
wherein a fully synthetic fundus image of the synthetic fundus used for generating the at least one synthetic fundus image is created by feeding input into a pre-trained latent-space based generative adversarial network (GAN) that converts the input into latent vectors within an intermediary latent space, and feeding the latent vectors into a second GAN trained to generate a photo-realistic synthetic image of a fundus according to the latent vectors.
|