Treatment planning for radiation therapy is a complex process, as there are many machine parameters to determine for a treatment. To decrease the required labour and improve the plan quality, auto-planning systems have been developed, which can automatically generate high-quality treatment plans. These plans still have to be checked to ensure their quality, which means an independent and automatic quality assurance method is needed. This is where deep learning comes in, as recent advancements in computer hardware and artificial intelligence make it possible to quickly train neural networks for fast predictions of dose distributions based on the anatomy of a certain patient. Even though these models have proven capable of generating predictions comparable to the actual delivered dose, there is no guarantee that the distributions are actually deliverable. Therefore, this study is conducted to develop a novel deep learning model that uses physical information of how dose is delivered to constrain the neural networks and force realistic and deliverable dose distributions. The models presented in this study are implemented on prostate cancer patients treated with the Volumetric-Modulated Arc Therapy (VMAT) delivery technique.

First, the conventional Anatomy-based dose prediction model is further developed to improve the quality of the predicted dose distributions. The goal is to assess how accurate the predicted dose is from a deep learning model without any physical constraint. The model with U-net architecture has shown to be capable of predicting dose distributions of very high quality with average DVH differences of less than 1 Gy and errors in the dose coverage statistics on the PTV of less than 1\%. Furthermore, the lack of deliverability was confirmed due to the missing ray effects in the dose distributions, showing no clear entry points of the external photon beam. This lack of ray effects was quantified and confirmed using the dip in the S\o rensen-Dice Index around the 30\% isodose contours, where the ray effects are most apparent.

Second, the Physics-guided prediction model was developed by combining the Anatomy-based dose prediction model with the newly developed Segment prediction model, which predicts Multi-Leaf Collimator (MLC) positions and beam intensity values for all directions from which the dose is delivered. These predictions are based on the prediction from the Anatomy-based dose prediction model and the Beam's Eye View images of the CT and the delineated structures. The segment prediction is used in a stand-alone dose engine, based on the Matrad dose calculation algorithm, to calculate a dose distribution from the predicted treatment plan, which is deliverable by definition. The performance of the Physics-guided prediction model is underwhelming, scoring considerably worse on most evaluation metrics compared to the Anatomy-based dose prediction model. The lack of performance is mainly caused by the poor performance of the Segment prediction model component.

To circumvent the issue of the Segment prediction model, a dose mimicking model has been developed and investigated, where a dose mimicking algorithm is used to mimic a prediction from the Anatomy-based dose prediction model by optimizing the segments on which the dose is delivered. This model has only been tested on a single patient but seems to be performing better than the Physics-guided prediction model, with the downside that an optimisation process is involved. Still, the performance is not on par with the Anatomy-based dose prediction model.

Further investigation on the Segment prediction model will be needed before the Physics-guided prediction model can be considered as a serious quality assessment tool. Once performance of the Segment prediction model has increased, it could prove an interesting model that could even replace an auto-planning system in the far future.

Radiotherapy treatment planning is a complex and time consuming process prone to differences as result of choices of individual planners. Autoplanning systems have been introduced to both reduce the time consumption and to counteract the influence of individual planning choices. Although autoplanning generally increases performance of the treatment plans, the plans still need to be checked manually to ensure plan accuracy. Increasing plan consistency with knowledge based planning (KBP) or automatic plan evaluation would be clear ways to improve upon this problem. For both improvements, deep learning can be an important tool to produce accurate and fast 3D dose distributions. In this study such a deep learning tool is developed in two parts to accommodate this.
In the first part a structure based deep learning model, using a U-Net architecture is developed, used for dose distribution prediction for prostate patients treated with volumetric modulated arc therapy (VMAT) plans, generated by autoplanning. Different loss functions have been tested to see which achieve the best results. In the second part physics information is included in the neural network prediction by using a hybrid physics-data approach. For this, an approximate dose distribution, the segment dose, is used as extra input for the dose prediction network. This segment dose is created by predicting multi leaf collimator (MLC) positions and reconstructing the corresponding dose with a simple dose engine. The predictions are evaluated using several dose characteristics, dose volume histogram (DVH) curves and other clinically relevant metrics. The U-Net architecture proved able to accurately predict the dose of patients in the test set. The best performing loss function for accurate dose characteristics prediction was found to be the weighted mean squared error (WMSE) loss, which predicted several PTV DVH statistics within a 1.5 % error and the rectum statistics within 3.5 % error. Furthermore, the model predicts the DVH points of the PTV within 0.84 ± 0.40 Gy average absolute dose difference and the rectum in 0.91 ± 0.72 Gy average dose difference. The hybrid physics-data model did not improve the prediction accuracy of the dose prediction model in terms of DVH statistics and DVH prediction, as the MLC positions could not be predicted accurately enough. The performance of the structure based prediction is similar to the performance of state-of-the-art prediction models. Although the hybrid model with the predicted segment doses did not improve the prediction accuracy, a significant effect could be seen from using the correct MLC positions instead. The bottleneck of the prediction is therefore identified to be the segment prediction. As such, it would be interesting to focus on segment prediction in follow up research.

Introduction:  Despite the vast amount of optimization algorithms, radiotherapy treatment planning remains a manual, time-consuming and iterative process. To increase plan standardization, we clinically use Pinnacle's autoplanner for several disease sites. However, this introduces new challenges: first, the autoplanner is not perfect and still requires substantial interaction from the radiotherapy technician (RTT). Second, it is difficult to judge whether a plan has indeed the most optimal trade-o_ between cure and toxicity, since the RTT has not worked the plan. Knowledge-based planning (KBP) could serve as a quality assurance tool to resolve these problems. It uses historical data (anatomical and dosimetric) from previous plans, to predict the likely dose distribution for the current patient. In this study, we construct an initial, simplistic KBP model that serves as the clinical practice. We then investigate of a variety of KBP modelling approaches to predict rectum dose-volume histograms (DVHs), in order to complement the current clinical practice in prostate cancer. Methods:  For model evaluation, we formulate a clinical tolerance criterion (TC) bandwidth based on a ground-truth set of existing radiotherapy plans. We evaluate on the overall prediction accuracy (RMS), the fraction of correctly predicted DVH bins (TC_α), anαd on the fraction of patients that have ≥ 90% of their DVH correctly predicted (TC_β). We use the overlap volume histogram (OVH) to encode for organ geometrical information, and use reduced order modelling (ROM) to capture the most important characteristics of the DVH and OVH. Optimization methods we use are Principal Component Analysis (PCA) eigenvalue RMS minimization, direct DVH RMS minimization, and TC_α and TC_β maximization. Results:  Analyses of the KBP clinical practice yielded training and testing errors of 81.4% and 80.8% for TC_α and 53.3% and 51.1% for TC_{β, with an RMS of 4.80 and 4.94 volume percentage [%]. Eigenvalue-optimization resulted TCα of 86.5% and 82.4%, and TCβ of 68.8% and 59.1%, with respective RMS of 2.82 % and 3.22 %. Direct DVH-optimization yielded TCα of 86.7% and 81.9%, and TCβ of 69.4% and 61.4%, with similar RMS. TCα and TCβ maximizers resulted TCα training and testing errors of 92.1% and 78.5%, and TCβ training and testing errors of and 84.3% and 53.4% respectively. Discussion:  The investigated models yielded significant improvements for direct eigenvalue- and DVH optimization methods. We have also been able to perform optimizations for the clinical goal metrics, showing promising results in training data. Because TCα- and TCβ- maximizers were unable generalize to perform well for unseen data, it is believed these metrics are too sensitive to be trained reliably, and more consistent data may be required for these optimizers to produce reliable test errors. Based on our findings, we advice the clinical practice to extend KBP-approaches to optimize for DVH-least squares.}