1. Multi-objective optimization as a novel weight-tuning strategy for deformable image registration applied to pre-operative partial-breast radiotherapy A first step toward uncovering the truth about weight tuning in deformable image registration 2. Kleopatra Pirpinia k.pirpinia@nki.nl +31 (0)20 512 2235 Dept. of Radiation Oncology Netherlands Cancer Institute (NKI) P.O. Box 90203 1006 BE Amsterdam The Netherlands Peter A.N. Bosman peter.bosman@cwi.nl +31 (0)20 592 4265 Centrum Wiskunde & Informatica (CWI) P.O. Box 94079 1090 GB Amsterdam The Netherlands Claudette E. Loo c.loo@nki.nl +31 (0)20 512 1088 Dept. of Radiology Netherlands Cancer Institute (NKI) P.O. Box 90203 1006 BE Amsterdam The Netherlands Astrid N. Scholten a.scholten@nki.nl +31 (0)20 512 2316 Dept. of Radiation Oncology Netherlands Cancer Institute (NKI) P.O. Box 90203 1006 BE Amsterdam The Netherlands Jan-Jakob Sonke j.sonke@nki.nl +31 (0)20 512 1723 Dept. of Radiation Oncology Netherlands Cancer Institute (NKI) P.O. Box 90203 1006 BE Amsterdam The Netherlands Marcel van Herk marcel.vanherk@manchester.ac.uk +44 (0) 161 918 2339 Institute of Cancer Sciences University of Manchester Wilmslow Road M20 4BX Manchester United Kingdom Tanja Alderliesten t.alderliesten@amc.uva.nl +31 (0)20 566 6886 Dept. of Radiation Oncology Academic Medical Center (AMC) P.O. Box 22660 1100 DD Amsterdam The Netherlands 3. Peter A.N. Bosman 4. Deformable image registration (DIR) has potential to enable novel approaches in radiotherapy (RT) such as dose accumulation, online adaptive planning, and response monitoring. Although DIR is predominantly formulated as a single-objective optimization problem, its inherent nature is multi-objective, i.e., there are multiple, conflicting objectives that need to be optimized simultaneously. A major challenge that limits its use in clinical practice, however, is the difficulty in choosing the optimal trade-off of these multiple objectives. Currently, primarily trial-and-error approaches are used to find weights to linearly combine multiple objectives into a single-objective function. Their success relies on a logical relation between the weights, objective values, and registration outcome, which is not well established. In this work, for the task of RT tumor response monitoring, we employ a multi-objective optimization approach that is not necessarily dependant on this logical relation and provides insightful tuning of weights even for hard registration cases. Deformable image registration is currently predominantly solved by optimizing a weighted linear combination of objectives. Successfully tuning the weights associated with these objectives is not trivial, leading to trial-and-error approaches. Such an approach assumes an intuitive interplay between weights, optimization objectives, and target registration errors. However, it is not known whether this always holds for existing registration methods. To investigate the interplay between weights, optimization objectives, and registration errors, we employ multi-objective optimization. Here, objectives of interest are optimized simultaneously, causing a set of multiple optimal solutions to exist, called the optimal Pareto front. Our medical application is in breast cancer and includes the challenging prone-supine registration problem. In total, we studied the interplay in three different ways. First, we ran many random linear combinations of objectives using the well-known registration software elastix. Second, since the optimization algorithms used in registration are typically of a local-search nature, final solutions may not always form a Pareto front. We therefore employed a multi-objective evolutionary algorithm that finds weights that correspond to registration outcomes that do form a Pareto front. Third, we examined how the interplay differs if a true multi-objective (i.e., weight-free) image registration method is used. Results indicate that a trial-and-error weight-adaptation approach can be successful for the easy prone to prone breast image registration case, due to the absence of many local optima. With increasing problem difficulty the use of more advanced approaches can be of value in finding and selecting the optimal registration outcomes. 5. G 6. Over the past few decades advances in computer science have led to image processing techniques that have proven useful for healthcare. One particular image processing task that can be of great value for healthcare, is deformable image registration: matching medical images, e.g., CT scans, taken at different times to explain what changed and how. An example of a potential application in clinical practice is matching a pre-operatively and post-operatively acquired CT scan from a breast cancer patient. Surgery is often only part of the treatment of breast cancer, being followed by radiotherapy. To deliver radiotherapy as best as possible (i.e., sparing as much healthy tissue as possible, yet hitting as much potentially still problematic tissue as possible), the area where the tumor was located prior to surgery, should receive a relatively high dose (a so-called boost). Post-surgery radiotherapy is planned on a CT scan acquired after surgery since this is most representative for the anatomy to be treated. However, on the CT scan acquired after surgery, the tumor is no longer visible and large deformations have occurred. It is therefore extremely hard for clinicians to make a reliable radiotherapy plan. Deformable image registration has the potential to solve this problem by identifying what went where and thus where the tumor was formerly located, making it clear how the radiotherapy plan should be designed so that the desired boost is realized. Over many decades, various highly efficient deformable image registration techniques have been developed. However, despite significant progress, deformable image registration is still not broadly used in clinical practice and challenging problems still remain. In particular, existing deformable image registration techniques compute the outcome of registration based on a single combination of different objectives. Typically, these objectives are related to the image similarity and the smoothness of the transformation. It is, however, an unsolved problem how to select the singular optimal combination of objectives beforehand, making clinical implementation of such algorithms difficult. Different combinations lead to different outcomes, which can ultimately only be judged in quality by experts. Hence, as also stated in our latest paper: a major challenge that limits its use in clinical practice, is the difficulty in choosing the optimal trade-off of these multiple objectives. Currently, primarily trial-and-error approaches are used to find weights to linearly combine multiple objectives into a single-objective function. Their success relies on a logical relation between the weights, objective values, and registration outcome, which is not well established. This problem is a well-known problem in the medical image processing community, yet no viable solution has appeared in the past that has the potential to bridge the gap between the potential of deformable image registration approaches and clinical practice. In other words, this setting meets the requirements of the HUMIES competition: the scientific community has long vetted this problem. With our approach, we believe a breakthrough has been achieved, paving the road for insightful, repeatable, high-quality deformable image registration results to be selected by medical experts, which can be used in clinical practice (i.e., the actual treatment of patients). As proof of this, the content of the papers itself serves as a basis, showing that the manual trial-and-error tuning problem is very hard, especially for clinically interesting problems such as the pre- and post-surgery case described above, or, as in the paper, a prone-supine breast image registration, which is also of high importance to clinical treatment. Furthermore, the partners in this project that this research is part of, are radiology, surgery, and radiotherapy departments and include clinicians to ensure that results obtained have real-world relevance and are not only statistically significant, but also clinically significant. We understand that this still falls within the circle of the authors themselves, which is insufficient as per the rules of the HUMIES competition. However, additional evidence comes from the scientific community (not the author, the author's own institution, or the author's close associates): for this work, the responsible PhD student (Kleopatra Pirpinia) has been nominated at the upcoming ICCR conference for the Young Investigator prize, showing acknowledgement from the problem-specific community (radiotherapy) rather than the EA community itself. Finally, although we believe that our results so far are exemplary and illustrate the idea and its impact already, results that are currently in preparation to be submitted for publication are even more exemplary to support our case (and will for instance include more data, even harder deformable image registration cases, and a direct comparison with true human expert performance rather than a surrogate thereof), we have decided to enter the competition THIS year because one of the co-authors is the general chair of GECCO 2017, which we feel would make participation for this award in that year uncalled for. 7. K. Pirpinia, P.A.N. Bosman, C.E. Loo, A.N. Scholten, J.-J. Sonke, M. van Herk and T. Alderliesten. Multi-objective optimization as a novel weight-tuning strategy for deformable image registration applied to pre-operative partial-breast radiotherapy. In Proceedings of the International Conference on the use of Computers in Radiation Therapy - ICCR-2016. (To Appear) K. Pirpinia, P.A.N. Bosman, J.-J. Sonke, M. van Herk and T. Alderliesten. A first step toward uncovering the truth about weight tuning in deformable image registration. In S. Ourselin and M.A. Styner, editors, Proceedings of the SPIE Medical Imaging Conference 2016. 978445; doi:10.1117/12.2216370, SPIE, Bellingham, WA, 2016. 8. We request that any prize will be divided equally among the 3 main institutes that have played a key role in this project: NKI, AMC, and CWI. For this, representative authors are Kleopatra Pirpinia of NKI (the PhD student doing the majority of the work), Tanja Alderliesten of AMC (daily supervisor, from the medical side, and original PI of the project that this work is a part of), and Peter Bosman of CWI (also daily supervisor, from the mathematics and computer-science side). 9. The new insights obtained by solving the medical deformable image registration problem with multi-objective EA approaches open new doors for an entire research field (deformable image registration). This particular research field has high potentials for improving clinical practice, which includes oncology (see also 6), but this potential has so far been realized only limitedly. With cancer being among the biggest health threats of society today, we believe that the societal relevance of our entry is huge. Furthermore, although a long-existing community of medical image processing researchers has achieved important results and has developed highly efficient software, the practical use of EAs that stems from the insight of EA researchers into this specific problem domain, provided an eye-opener that the medical image processing community hadn't yet discovered for itself. This shows that EA research(ers) and software cannot only provide novel human-competitive results, but also create novel, added value for an entire community of human researchers, which may be considered a human-competitive result in itself. 10. The general type of evolutionary computation used, is Estimation-of-Distribution Algorithms (EDAs), which are nowadays considered to be a specific type of Model-Based Evolutionary Algorithms. In particular, the EDAs used are real-valued and bear similarities with several Evolution Strategies (ES).