Oxı̇

@xgranade@wandering.shop If you only use synthetic data a trained net can only register the synthetic data. if you fine tune the trained net with real images the results are better than training only using the real data (<100 image pairs)

@xgranade@wandering.shop This is more a very niche case and not a counter example for your theory

@xgranade@wandering.shop CycleGANs might be usefull. A colleague of mine used it to generate MRI and CT images from a digital phantom so we had data to train and evaluate multimodal image registration.