ESTRO 38 Abstract book

S117 ESTRO 38

- Personnel costs and factual investments within the biomarker-based procedures. The main cost drivers for radiation-drug combinations will probably be: - The medication and management of adverse events for combination A - The diagnostic procedures for combination B

SP-0235 Intuitive and Insightful Evolutionary Intelligent Treatment Planning P. Bosman 1 1 Centrum Wiskunde and Informatica CWI, Life Sciences and Health, Amsterdam, The Netherlands Abstract text Artificial Intelligence (AI) increasingly pervades the daily news, with self-driving cars, robots, and face recognition even on smart phones. The reason for this revolution is threefold: 1) several AI techniques have matured algorithmically, 2) there have been spectacular advances in computing hardware (e.g., GPUs), and 3) digitizing and storing large amounts of data has become common practice. AI will increasingly be a key driver of automation, including in radiation oncology. In this talk, placed within the broader context of existing automation approaches to brachytherapy treatment planning in general, I will present the AI efforts undertaken so far by my research group to automate. These are mainly focused around a particular subfield of AI: that of Evolutionary Algorithms (EAs). EAs are loosely based on the natural evolution principles and have the advantage of being relatively straightforward to apply to optimization and machine learning tasks. A disadvantage is that most textbook EAs are ``blind'', as they randomly combine solutions. I will introduce the state-of-the-art Gene-pool Optimal Mixing Evolutionary Algorithm (GOMEA) family that automatically determines how to best combine solutions. I will illustrate the advantageous properties of GOMEAs over classical ``blind'' EAs, showing how GOMEAs are capable of obtaining (near-)optimal results for problems with millions of variables in less than an hour on a normal desktop computer, where classic EAs can only do this for tens or a few hundred variables. I will show how with GOMEA we are able to solve bi- objective problem formulations based directly on dose- volume indices for automated HDR brachytherapy treatment planning for the prostate, thereby finding not just one treatment plan, but trade-off curves consisting of many interesting potential treatment plans that trade-off target coverage with sparing organs at risk for large numbers of dose calculations points in mere minutes on a single modern GPU. Visualizing such a trade-off curve gives an insightful overview of what types of dose-volume indices are achievable for a particular patient. Beyond this, I will briefly touch upon near-future avenues for further AI-based automation. SP-0236 MV reference dosimetry in TRS-398: State-of- the art and research supporting an updated code of practice F. Delaunay 1 , C. Andersen 2 , L. De Prez 3 , S. Duane 4 , M. Pimpinella 5 , P. Teles 6 , J. Tikkanen 7 , K. Zink 8 1 CEA- List, Laboratoire National Henri Becquerel Lne- Lnhb, F-91191 Gif-Sur-Yvette, France ; 2 DTU, Center For Nuclear Technologies, Kongens Lyngby, Denmark ; 3 VSL, Dutch Metrology Institute, Delft, The Netherlands ; 4 national Physical Laboratory, Chemical- Medical and Environmental Science Department, Teddington, United Kingdom ; 5 ENEA, National Institute Of Ionizing Radiation Metrology Inmri, Rome, Italy; 6 University of Lisbon, GPSR- Centre of nuclear sciences and technology- IST, Lisbon, Portugal ; 7 Stuk, Radiation And Nuclear Safety Authority, Helsinki, Finland; 8 University Of Applied Symposium: Reference and non-reference dosimetry - CoPs and beyond

Symposium: Inverse planning in brachytherapy - A one click solution?

SP-0233 Optimal use of inverse optimization in brachytherapy D.Baltas 1 1 Medical Center - University of Freiburg, Division of Medical Physics - Department of Radiation Oncology, Freiburg, Germany

Abstract not received

SP-0234 Inverse treatment planning in clinical practice, one click and done? D. Todor 1 1 Virginia Commonwealth University, Radiation Oncology, Richmond- VA, USA Abstract text Treatment Plan Optimization is one of those subjects which strongly polarizes brachytherapy practitioners. Part of the problem is that we got where we are today by slowly building on decades of successful clinical experience, relying on clinical knowledge sublimated in collections of rules called ‘systems’: Stockholm, Paris, Manchester. These systems described, sometimes in great details, how things should be done; they represent practice not theory. This approach had a lot to do with reproducibility of treatments that seem clinically successful and little to do with understanding how things work, why a certain placement of radioisotopes actually works and explaining the mechanism behind. Fast forward to the current era in which we learned how to compute radiation transport through the body very accurately and to describe our plans by employing dose volume histograms (DVH). Having now a set of DVH parameters that describes a dose distribution, one is in the position of optimizing that distribution by simultaneously maximizing those parameters to some structures (anatomical or disease related) while minimizing other parameters to structures related to normal anatomy, which typically one tries to spare. Optimization is an interesting applied mathematical domain and attracts computer scientists, AI researchers, physicists and mathematicians. The less than optimal reality (pun intended) is that optimization is seldom used in the clinic (at least in the brachytherapy practice) and when it is used, a physician or a physicist will typically ‘adjust’ a solution obtained straight out of the optimizer with the intent of further improving it. My plan is to discuss the origins and the [i]rationale of this state of affairs as well as to describe possible avenues for increased utilization of true optimization. I will present three cases (cervical, prostate and breast cancer) as how- to examples of what optimization can deliver. I will devote a second part of my talk describing the obvious fact that plan optimization really is dose optimization and more precisely DVH-based optimization. I will do my best to question central concepts like prescription dose, targets, dose coverage, Ds and Vs, (all ingredients in the Optimization) and try to setup alternative constructs. An example based on cervical cancer data and/or prostate cancer will follow.