2 nd ESTRO Forum 2013 S111 have benefitted greatly from these interactions.Results of our collaboration can be seen in the ROSIS project and associated short courses, involvement in clinical audit as members of the multidisciplinary team and several joint publications.It is a great honour to receive this award and I am delighted to 'finally' be a physicist albeit it an honorary one -but maybe that is even better!

In order to provide a consistently high quality treatment, it is of great interest to assess the robustness of a treatment plan under the influence of geometric uncertainties. One possible method to implement this is to run treatment simulations for all scenarios that may arise from these uncertainties. These simulations may be evaluated in terms of the statistical distribution of the outcomes (as given by various dosimetric quality metrics) or statistical moments thereof, e.g. mean and/or variance. This paper introduces a method to compute the outcome distribution and all associated values of interest in a very efficient manner. This is accomplished by substituting the original patient model with a surrogate provided by a machine learning algorithm. This Gaussian process (GP) is trained to mimic the behavior of the patient model based on only very few samples. Once trained, the GP surrogate takes the place of the patient model in all subsequent calculations.The approach is demonstrated on two examples. The achieved computational speedup is more than one order of magnitude.

We present a method of modeling dosimetric consequences of organ deformation and correlated motion of adjacent organ structures in radiotherapy. Based on a few organ geometry samples and the respective deformation fields as determined by deformable registration, principal component analysis (PCA) is used to create a low-dimensional parametric statistical organ deformation model (Söhn et al 2005 Phys. Med. Biol. 50 5893-908). PCA determines the most important geometric variability in terms of eigenmodes, which represent 3D vector fields of correlated organ deformations around the mean geometry. Weighted sums of a few dominating eigenmodes can be used to simulate synthetic geometries, which are statistically meaningful inter- and extrapolations of the input geometries, and predict their probability of occurrence. We present the use of PCA as a versatile treatment simulation tool, which allows comprehensive dosimetric assessment of the detrimental effects that deformable geometric uncertainties can have on a planned dose distribution. For this, a set of random synthetic geometries is generated by a PCA model for each simulated treatment course, and the dose of a given treatment plan is accumulated in the moving tissue elements via dose warping. This enables the calculation of average voxel doses, local dose variability, dose-volume histogram uncertainties, marginal as well as joint probability distributions of organ equivalent uniform doses and thus of TCP and NTCP, and other dosimetric and biologic endpoints. The method is applied to the example of deformable motion of prostate/bladder/rectum in prostate IMRT. Applications include dosimetric assessment of the adequacy of margin recipes, adaptation schemes, etc, as well as prospective 'virtual' evaluation of the possible benefits of new radiotherapy schemes.

Frequently, radiotherapy treatments are comprised of several dose distributions computed or optimized in different patient geometries. Therefore, the need arises to compute the comprehensive biological effect or physical figure of merit of the combined dose of a number of distinct geometry instances. For that purpose the dose is typically accumulated in a reference geometry through deformation fields obtained from deformable image registration. However, it is difficult to establish precise voxel-by-voxel relationships between different anatomical images in many cases. In this work, the mathematical properties of commonly used score functions are exploited to derive an upper boundary for the maximum effect for normal tissue and a lower boundary for the minimum effect for the target of accumulated doses on multiple geometry instances.

PURPOSE:Organ movement is still the biggest challenge in cancer treatment despite advances in online imaging. Due to the resulting geometric uncertainties, the delivered dose cannot be predicted precisely at treatment planning time. Consequently, all associated dose metrics (e.g., EUD and maxDose) are random variables with a patient-specific probability distribution. The method that the authors propose makes these distributions the basis of the optimization and evaluation process.METHODS:The authors start from a model of motion derived from patient-specific imaging. On a multitude of geometry instances sampled from this model, a dose metric is evaluated. The resulting pdf of this dose metric is termed outcome distribution. The approach optimizes the shape of the outcome distribution based on its mean and variance. This is in contrast to the conventional optimization of a nominal value (e.g., PTV EUD) computed on a single geometry instance. The mean and variance allow for an estimate of the expected treatment outcome along with the residual uncertainty. Besides being applicable to the target, the proposed method also seamlessly includes the organs at risk (OARs).RESULTS:The likelihood that a given value of a metric is reached in the treatment is predicted quantitatively. This information reveals potential hazards that may occur during the course of the treatment, thus helping the expert to find the right balance between the risk of insufficient normal tissue sparing and the risk of insufficient tumor control. By feeding this information to the optimizer, outcome distributions can be obtained where the probability of exceeding a given OAR maximum and that of falling short of a given target goal can be minimized simultaneously.CONCLUSIONS:The method is applicable to any source of residual motion uncertainty in treatment delivery. Any model that quantifies organ movement and deformation in terms of probability distributions can be used as basis for the algorithm. Thus, it can generate dose distributions that are robust against interfraction and intrafraction motion alike, effectively removing the need for indiscriminate safety margins.

This article reports on a 4D-treatment planning workshop (4DTPW), held on 7-8 December 2009 at the Paul Scherrer Institut (PSI) in Villigen, Switzerland. The participants were all members of institutions actively involved in particle therapy delivery and research. The purpose of the 4DTPW was to discuss current approaches, challenges, and future research directions in 4D-treatment planning in the context of actively scanned particle radiotherapy. Key aspects were addressed in plenary sessions, in which leaders of the field summarized the state-of-the-art. Each plenary session was followed by an extensive discussion. As a result, this article presents a summary of recommendations for the treatment of mobile targets (intrafractional changes) with actively scanned particles and a list of requirements to elaborate and apply these guidelines clinically.

Purpose: Organ movement is still the biggest challenge in prostate treatment despite advances in online imaging. Special robust optimization techniques produce organ doses that are insensitive against organ movement. Robust optimization requires a statistical patient model. Based on a finite number of CTs, the movement of the organs is estimated quantitatively. We investigate the interplay of patient model and robust optimization technique. In particular, the minimum number of images necessary to obtain a dependable robust treatment plan is determined. Materials: Starting from N CT images, a statistical shape model of the patient is created by Principle Component Analysis. This statistical information is incorporated into the robust treatment plan optimization. Organ motion gives rise to uncertainty in the treatment outcome parameters (i.e. EUD, etc.). By propagating the organ geometry uncertainty all the way into the treatment outcome parameters, we are able to predict and shape the outcome distributions. This is in contrast to the conventional optimization of a nominal value. Note that the uncertainty in the dose distribution does not necessarily correlate with the uncertainty in treatment outcomes.The basis for our analysis were the CT datasets of four prostate cancer patients, consisting of ~15 images each. For each patient, N CTs were drawn at random from the pool. The patient model based on these CTs was used for optimization. The procedure was rerun 20 times. The resulting dose distributions, along with their respective treatment outcome distributions were analyzed for their similarity.Additionally, it was investigated whether the treatment outcome distributions as predicted by the optimizer coincide with the outcome distributions that are obtained if the model of motion is built using all available imagery, which serves as a gold standard.Both procedures were repeated for varying N. Results: Our analysis indicates that ~5 images suffice to generate a patient model that is able to capture all significant aspects of the patient’s movement. The exact number of required CTs varies from patient to patient, depending on the degree of movement. This suggests an adaptive radiotherapy scheme: a conventional treatment is launched and when enough CBCT images become available, the treatment outcome distributions are calculated. Based on these distributions the plan in place can either be verified or substituted with our robust approach. Conclusions: Our study shows that fully featured statistical optimization is possible under clinical conditions. PCA in conjunction with our optimization technique produces robust treatment plans based on a realistic amount of CTs. Given the small number of required CTs, the outcome distributions can be computed at an early stage of treatment. This allows for the verification as well as correction of existing plans based on probabilistic information.

The major challenge in intensity-modulated radiotherapy planning is to find the right balance between tumor control and normal tissue sparing. The most desirable solution is never physically feasible, and a compromise has to be found. One possible way to approach this problem is constrained optimization. In this context, it is worthwhile to quantitatively predict the impact of adjustments of the constraints on the optimum dose distribution. This has been dealt with in regard to cost functions in a previous paper. The aim of the present paper is to introduce spatial resolution to this formalism. Our method reveals the active constraints in a target subvolume that was previously selected by the practitioner for its insufficient dose. This is useful if a multitude of constraints can be the cause of a cold spot. The response of the optimal dose distribution to an adjustment of constraints (perturbation) is predicted. We conclude with a clinical example.