At an earlier stage of this project, yet another idea which was based on the relationship between registration and models, got investigated in practice. Sparked by successful work on shape models, wherein good models were constructed by minimisation of their complexity, a similar approach was applied to combined models, i.e. models incorporating both shape and intensity. Images were transformed in a well-controlled manner, their model was built and transformations were subsequently accepted or rejected based on the quality of the model.
Despite the impracticality of the algorithm, which suffered from heavy computational requirement, proof-of-concept results were quite satisfactory.