Deformation equivariant cross-modality image synthesis with paired non-aligned training data