Figure 1.

Our proposed self-supervised learning concept for multimodal image registration aiming to minimise a cycle discrepancy. In every training iteration, another (known) random transformation matrix R₂₃ is used to generate a synthetic image. Like this, a cycle consisting of two unknown multimodal transformations (with the transformation matrices R₂₁ and R₃₁) and a known monomodal transformation (with the transformation matrix R₃₁) is obtained, leading to the minimisation problem of $|{\mathrm{R}}_{23}·{\mathrm{R}}_{31}-{\mathrm{R}}_{21}|\to \min$ that is used for learning.