Exploring the potential and feasibility of multi-objective deformable image registration for breast cancer treatment

Deformable image registration (DIR), the process of deforming one image to match another, is an image processing technique of increasing importance in cancer diagnosis, monitoring and treatment, as it allows to, e.g., align image data acquired in different patient orientation, or capture patient anatomy changes over time. To solve the DIR problem, current DIR methods rely on multiple parameters that need to be tuned, a non-trivial task; further, it is difficult for these methods to capture large deformations and content mismatch present in the images. In part due to such obstacles, the use of DIR in clinical practice remains limited. In this thesis, to improve and facilitate use of DIR in clinical practice, we proposed a novel perspective on DIR based on multi-objective optimization, i.e., we did not model DIR as a problem with one unique solution, but as a problem with a set of solutions that can be considered equally good, each one representing a high-quality trade-off between DIR objectives of interest. We considered two challenging DIR problems in the context of breast cancer treatment, which involve large deformations and content mismatch: prone-to-supine breast MRI registration, and pre- to post-operative breast CT registration. In Part I, we saw that a multi-objective perspective on DIR can improve the way current DIR methods are fine-tuned, thereby improving their use in clinical practice for problems with limited deformations. In Part II, we made the first algorithmic steps towards developing a purely multi-objective DIR method, for solving hard registration problems.