The paper mentioned how the images were processed (chopping 10% off some to remove name, age, etc). But all were from the same centre and only pixel data was used. Given the other work referenced on retinal thinning in ASD disorders, maybe it is a relatively simple task for this kind of model. But they do say using multi-centre images will be an important part of the validation. It’s quite possible the performance would drop away when differences in camera, etc. are factored in.