ESTRO 2021 Abstract Book

S453

ESTRO 2021

Automatic treatment planning has seen a significant boost in recent years. The well-known causes are the continuous integration of well-established machine learning tools into commercial planning systems, together with the spectacular scientific production of state-of-the-art deep learning models. Treatment planning automation is a complex task, which in fact needs to integrate three different optimization loops. The first and best studied is to provide an optimal plan for given DVH objectives and constraints. This problem is now efficiently solved by most optimizers. However, the quality of the obtained dose distributions does not necessarily correspond to the clinical optimum. It therefore follows a relatively tedious process of manual optimization of the objectives, constraints, and penalties in order to guide the optimization of treatment towards the clinical optimum. This (second) step is usually performed by an expert, either a dosimetrist or a medical physicist. Finally, the last level is the evaluation by the radiation oncologist responsible for the treatment plan, who will provide final remarks to further update the optimization parameters or accept the plan if everything is as desired. At a minimum, the automation of treatment planning should address the last two levels. This generally results in the automatic generation of dose-volume objectives allowing the optimizer to provide the desired treatment plan and dose distribution. The first commercial and academic solutions generally focused on these tasks. A theoretical limitation of this approach is that all learning is done on data of reduced dimensionality (DVHs), which only partially represents the problem to be solved (finding the optimal 3D dose distribution)The advent of deep learning in the field of treatment planning, has changed the perspective. With deep learning, it is possible to directly predict in three dimensions the dose distribution from CT images and contours, without going through a treatment plan optimizer. In principle, this makes it possible to capture all the information available. However, these methods still have to use a conventional treatment plan optimizer in order to optimize a deliverable treatment plan whose associated dose distribution should closely reproduce the predicted dose distribution. During the presentation, we will present and compare the performance of the available methods. In addition, deep learning opens the possibility of generating clinical dose distributions without a treatment plan optimizer, and therefore without the hitherto essential support that is the treatment planning system. This gives rise to a whole series of new applications in the field of decision support for the choice of treatment modality, but also in the training of dosimetrists and medical physicists. All the solutions discussed so far are based on models trained or fitted using retrospective data. The need for data is important, both in quantity and in quality, especially for deep learning. Obtaining homogeneous databases of sufficient size representing the state of the art at a precise moment is not a trivial task. Expert planners can change, expertise can change, practice can change. We will show that the homogeneity of the data is of capital importance and that sometimes the importance of the quantitative aspect is overestimated. A final limitation due to the use of retrospective data, this time much more fundamental, is the impossibility of improving clinical practice beyond reducing variability. In other words, the current automation tools can only reproduce and systematize an existing clinical approach, but cannot improve it or propose novel ones. In this presentation, we will present what are the current avenues that could allow a paradigm shift where artificial intelligence could generate knowledge instead of simply reproducing it. Abstract Text Through many developments over the past decades, the quality and therapeutic effect of radiotherapy has greatly improved. Dose distributions can be designed much more precise than before, there are multiple treatment modalities as option, and there is a request for further personalisation. However, these will increase the workload in clinics. Also, there are already many people involved in initiating a radiotherapy treatment: physicians, technicians, physicists, who are all active in different part of the process, and require optimal communication. An ideal situation would be where the “treating physician is in control” over the full delineation and treatment planning process, where all phases in the process can be performed by one person directly and in real-time. Thus right after finishing (or checking) the delineations, the physician can actively explore different treatment options, or choose for more personalised trade-offs. One of the other advantages is that all actions lie with one person, avoiding communication and transfer of treatment ideas with the technician, and that the physician can focus on a patient in one unfragmented time-window. While automated treatment planning enables plan generation for a large range of treatment options and also avoids increasing the planning workload, the mathematical optimisation process is too lengthy. Processing the large amount of data required for optimisation simply takes time, also on modern fast computer hardware. There is thus a hard time limit for mathematical generation of accurate and optimal treatment plans. Deep learning offers an alternative: it has been demonstrated that deep learning has the potential for dose prediction in several seconds (nearly instantaneous), which would enable the ideal situation for the new workflow. However, most publications have been proof-of-concepts, and the clinical quality and applicability has not yet been thoroughly assessed. An attribute of deep learning methods is that they require vast amounts of high quality and consistent data to be trained well. Existing clinical treatment plans are not very consistent, have variable quality, and are only a single plan per patient (i.e. do not offer multiple treatment options). This is where automated treatment planning through mathematical optimisation meets deep learning. Automated treatment planning generates high quality and consistently planned clinically preferable treatment plans, while with modified configurations, can also generate many alternative treatment options and plans for different modalities for the same patient. As a basis, only a delineated CT is required, of which large amounts are readily available in the clinical archives. As such, automated treatment planning is capable of generating sufficient data to train deep learning models. It is still to be investigated how well and how consistent deep learning performs in treatment planning, and SP-0580 Deep learning and/or mathematical optimisation for automated planning S. Breedveld 1 1 Erasmus MC, Radiotherapy, Rotterdam, The Netherlands