Background and purpose: The variable RBE of protons could be strategically harnessed by confining high RBE to the tumor by optimizing the LET distribution, which has been shown to strongly correlate with RBE and does not contain the uncertainties associated with variable RBE models. However, increasing target LET compromises target dose robustness against setup and range uncertainties, as LET rises sharply past the Bragg peak, where the dose gradient is steep. Previous studies can lack thorough robustness analyses and rely on LET-based objectives in optimization, which are not available to all institutions. Lung cancer, in particular, remains underexplored in the context of target-LET optimization. This study therefore (i) investigates the elevation of target LET without relying on LET-based optimization objectives, (ii) determines the maximum feasible increase while maintaining
adequate target dose robustness and adherence to clinical goals, and (iii) assesses whether such increases translate into meaningful improvements in tumor control probability in a cohort of recently treated lung cancer patients.
Materials and methods: A LET-painting technique using two sets of opposing beams with beam-specific minimum dose objectives for proximal target segments was employed to elevate target LET. To find the near-feasible target LET, high inhomogeneity was
leveraged in combination with CTV-based planning, under the assumption of motion-mitigation. Probabilistic evaluation using Polynomial Chaos Expansion (PCE) was then performed to ensure precise adherence to adequate robustness. LET-optimized
treatment plans, employing this recipe, were then compared to dose-optimized treatment plans for 10 lung cancer patients, based on dose-averaged LET (LET_d), LET_d×D, Unkelbach RBE, and EUD-based TCP values.
Results: An increase of 50% in target LET_d was observed for the probabilistically robust LET-optimized treatment plans compared to the dose-optimized plans, raising the RBE-weighted dose by 6 Gy and mean target RBE from 1.1 to 1.16. These gains were
associated with increased mean target dose, reduced target dose robustness and increased beam-entry doses. The elevated LET_d significantly improved tumor control probability, although a large part of the total gain in TCP was due to increased physical dose.
Conclusion: Clinically significant increases in TCP due to enhanced target LET_d are possible without the use of LET_d-based objectives, by leveraging high inhomogeneity and precisely adhering to adequate target dose robustness. It should be determined
whether these trade-offs in addition to changes in OAR doses are clinically acceptable and should be directly compared to dose escalation strategies alone. Future technologies might help enhance benefits or improve trade-offs between inhomogeneity, robustness
and target LET, making target RBE-enhancement through LET to improve treatment efficacy more clinically viable. ... Read more

Evidence indicates that ultra-high dose rate (UHDR) irradiation in radiotherapy can induce a normal tissue sparing effect without compromising effectiveness against tumour cells, known as the FLASH effect. This has prompted active research into clinical proton FLASH therapy. A key step is developing a clinically safe and predictable proton FLASH beam, which is being pursued through the commissioning of the ProBeam gantry in FLASH mode at HollandPTC. As part of this process, this thesis aims to characterise the gantry-based 250 MeV ultra-high dose rate continuous scanning proton beam, currently only intended for preclinical research.

Four distinct dosimetric aspects have been investigated, in accordance with the AAPM TG-224 report and machine quality assurance guidelines for UHDR proton beams in transmission mode. The first three, which are also part of conventional characterisations, include lateral and longitudinal relative dosimetry, along with absolute dosimetry measurements. These encompass the spot shape, spot position, integral depth dose (IDD) curve and output measurements. The fourth category includes temporal dosimetry, an essential aspect for FLASH characterisations since delivery time has now become an important aspect. For this, the scanning speed, spot dwell time and dose rate (constancy) have been determined. The temporal measurements were conducted using the FlashQ detector, a 2D strip ionisation chamber with a temporal resolution of 1 ms, which also served as a reference monitor chamber to enable the full spatiotemporal reconstruction of each irradiation. Measurements were conducted at nominal nozzle currents from 2 to 215 nA to explore potential correlations between the measured parameters and nozzle current. The suitability of all other detectors used in this work for FLASH measurements has also been assessed.

The spot shape did not show a clinically significant dependency on the nozzle current, and the average Gaussian parameters were determined to be σx = 3.39 ± 0.06 mm, σy = 3.86 ± 0.02 mm and θ = -16.6 ± 4.5 degrees. Spot position accuracy was within 0.15 mm, complying with AAPM TG-224 standards. The R80-value from the integral depth dose (IDD) measured 37.9 ± 0.1 cm, aligning with IDD results from other ProBeam facilities. Absolute dose per monitor unit (MU) varied significantly with nozzle current, from 0.0041 Gy/MU at 10 nA to 0.0466 Gy/MU at 215 nA. In terms of temporal aspects, the gantry scanning speed was found to depend on the spot spacing but converged to 7.8 ± 0.1 m/s and 29.4 ± 0.9 m/s in the x- and y-directions respectively for large spot spacings (> 50 mm). The nozzle monitor chamber reached saturation at a nominal nozzle current of 18.4 nA, resulting in fixed spot dwell times. This saturation caused significant dose rate fluctuations, both day-to-day and beam-to-beam. Over six measurement sessions in a four-month period, deviations ranged from -13.9% to -23.2% compared to planned dose rates, with an average intraday fluctuation of 3.9%. These fluctuations were measured with the FlashQ, which has been verified as a suitable reference detector.

Ultimately, the gantry-based 250 MeV UHDR continuous scanning proton beam has been successfully characterised at HollandPTC. All conventional parameters met the AAPM TG-224 standards or aligned with findings from other ProBeam institutes. Using the FlashQ as a reference monitor chamber enables the reconstruction and simulation of dose delivery in both space and time through calibration.
... Read more

This thesis outlines the project conducted at HollandPTC for the Master Biomedical Engineering, track Medical Physics, at Delft University of Technology.

Metal implants create dosimetric uncertainties in treatment planning and delivery for proton therapy. The purpose of this thesis was to determine the most suitable treatment planning strategy for proton therapy in the presence of cranial fixation clips.
A phantom was designed and imaged using computed tomography. Subsequently, base experiments were performed on a mono-energetic proton beamline to investigate the dose perturbation caused by the clip. It was found that the presence of the clip in the radiation field produces a measurable signal perturbation and that the clip affects the field uniformity.

A total of seven different treatment planning strategies were created. These strategies were based on four main concepts that were determined from a literature review and were identified to have potential effectiveness in addressing cranial fixation clips for brain cancer: implementing a density override on the implant, employing either a single-field optimization (SFO) or multi-field optimization (MFO) technique, implementing beam-specific margins or not, and adding an avoidance margin on the implant.
All strategies, except for the one including an avoidance margin around the implant, were successfully implemented into the treatment planning system. The plans were robustly optimized and met the clinical goals regarding clinical target volume (CTV) coverage and organ at risk (OAR) sparing.

The strategies were evaluated by performing experiments on a clinical treatment gantry. Based on the results of the first set of experiments, a Lynx detector was used, which was currently not able to detect the measured signal in dose in Gray. The aim of the experiments was to gain insights into the agreement between the data from the treatment planning system and the experiments, as well as to investigate what the best strategy is for handling cranial fixation clips in proton therapy.

Based on the results of the conducted experiments, there is an indication that the field homogeneity inside the target volume gets perturbed by the clip. This indicates a high need to investigate the clinical implications of this further. In particular, absolute dose measurements must be added to the presented results.
... Read more

This thesis addresses two major challenges in modern radiotherapy workflows: the slow computation speed of dose prediction algorithms and the insufficient modeling of anatomical variations during and between treatment fractions. Current photon and proton therapy plans rely on pre-treatment computed tomography (CT) scans obtained days before the start of treatment. Inter-fraction anatomical changes, intra-fraction organ motion, and setup errors compromise treatment accuracy and may unnecessarily irradiate healthy tissue. Existing mitigation strategies—such as target margins in photon therapy and robust optimization in proton therapy—only partially address these uncertainties and are limited by the lack of realistic anatomical models and fast dose prediction methods.

The first part of this work presents millisecond-scale dose prediction algorithms for proton pencil beams and photon beams using deep learning. Chapter 2 introduces the Dose Transformer Algorithm (DoTA), a model that predicts proton beamlet doses by combining convolutional neural networks with a transformer backbone that captures both spatial features and beam energy information. DoTA achieves gamma pass rates above 99% while reducing computation time by four orders of magnitude compared to Monte Carlo simulations. Chapter 3 extends this approach to photons with the improved Dose Transformer Algorithm (iDoTA), which maps projected beam geometries to 3D dose distributions. iDoTA estimates full VMAT dose distributions in seconds with state-of-the-art accuracy, significantly accelerating conventional photon treatment planning.

The second part focuses on anatomical variations. Chapter 4 presents the Daily Anatomy Model (DAM), a probabilistic deep learning framework that generates patient-specific inter-fraction deformations of planning CT images based on population data. DAM captures correlated movements with few latent variables, accurately reproducing prostate volume and center-of-mass variations observed in repeat CT scans, and enabling robust treatment planning against daily anatomical changes. Chapter 5 models intra-fraction respiratory motion using variational and adversarial autoencoders, including a semi-supervised extension for joint signal classification and generation. A novel time-series compression method reduces multi-dimensional breathing cycles to low-dimensional vectors while preserving high-resolution reconstruction. These models generate realistic, class-specific breathing signals, supporting simulation of target motion during radiation delivery.

Chapter 6 applies these anatomical models to simulate interplay effects in Intensity Modulated Proton Therapy (IMPT), arising from interactions between tumor motion and scanning beam movement. Using both simple sinusoidal and deep learning-generated breathing signals, the analysis quantifies how small variations in respiratory period affect local dose distributions. The results highlight that conventional planning approaches, including 4DCT and Internal Target Volume (ITV) plans, often fail to achieve clinically required robustness, underscoring the need for individualized modeling.

In conclusion, this thesis provides methods to predict dose deposition with millisecond speed and simulate realistic anatomical variations for both inter- and intra-fraction motion. These contributions enable more accurate robustness evaluation, support future online adaptive workflows, and offer a foundation for integrating deep learning-based dose and anatomy models into clinical radiotherapy. Future research should focus on coupling these algorithms with existing treatment planning systems and validating their performance in diverse clinical scenarios. ... Read more

Heart failure is a leading cause of death and forms a growing health concern. The development of novel drugs is however hampered by the absence of adequate screening methods and disease models. Cardiomyocytes derived from patients could assist in the development of a patient specific drug screen method to test the efficacy and safety of putative drugs. Simultaneously, deep learning has been applied to a variety of biomedical datasets, achieving state-of-the-art performance. Previous methods for the classification of cardiomyocytes as healthy or diseased only focused on machine learning methods. We present the first deep learning approach to perform this classification task together with a novel artificial intelligence interpretability method called Contraction Analysis Local Interpretable model-agnostic explanations (CA-LIME), able to explain the predictions made by the classifier. The proposed classifier is shown to outperform previously developed methods to classify cardiomyocytes, obtaining 97.5% accuracy. Our results indicate this classifier could aid in the development of a high throughput drug screening system for cardiac drug development. The explanations made by CA-LIME are in correspondence with previous observations of drugs with known effects, verifying the effectiveness of our approach. Together with CA-LIME, the processing pipeline could lead to the discovery of new differences between the motion of healthy and aberrant beating cardiomyocytes.
... Read more

With almost 900.000 new patients per year suffering from head and neck squamous cell carcinomas (HNSCC), who after radiotherapy treatments experience severe side effects, the focus is drawn to proton therapy. Proton therapy results to less side effects since healthy tissues surrounding the tumour receive less radiation dose. This results from the proton’s specific depth-dose profile with its Bragg peak.
To further reduce the effect on surrounding tissues, major improvements should be made in understanding the biologic response to proton therapy which will eventually lead to enlarging the therapeutic window. Here in vitro analysis with DNA damage repair inhibition or increasing proton dose rates can be performed to assess which DNA repair pathways repair radiation induced DNA damage and to assess the effect of proton dose rates on biological tissues. This will be investigated in vitro, while in vivo experiments are necessary as well. However, currently most mice are irradiated with protons in the plateau of the Bragg curve instead of the Bragg peak. Therefore, a dosimetric pipeline should be designed to allow for accurate placement of mouse tumours in the proton’s Bragg peak.
In this thesis, the osteosarcoma- and HNSCC-derived U2OS and FaDu cell lines were incubated with non-homologous end-joining or homologous recombination inhibitors during x-ray and proton irradiations. We have shown a radiosensitising effect of the non-homologous end joining inhibitor AZD7648 on FaDu cells during x-ray irradiation, but an effect during proton irradiation could not be proved nor neglected. The homologous recombination inhibitor B02 could not effectively inhibit homologous recombination and was shown to not affect clonogenic survival of FaDu during X-ray irradiations. The observed radiosensitising effect of the non-homologous recombination inhibitor AZD7648 can be exploited in patient selection based on already existing DNA damage repair deficiencies in the tumour. Furthermore, combination therapies could be used of photon irradiations with AZD7648 targeted to the tumour to artificially enlarge the therapeutic window.
Besides this, the HNSCC-derived FaDu cells were irradiated with varying dose rates to assess induction of a FLASH effect. This FLASH effect is usually observed after irradiations of >40 Gy/s resulting in a reduction of normal tissue complications while tumour control is maintained. In our analysis, small differences in low dose rates did not have any impact. The experimental set up for proton FLASH irradiations of cells was prepared and though we performed the first set of experiments no definite conclusion could be drawn due to the level of biological variation we observed.
Of course, differential effects between tumours and tissues like the FLASH effect should be investigated in vivo. Therefore, a set-up should be created to irradiate the mural tumour with Bragg peak protons. To reach this, two micro-CT scanners were calibrated to link Hounsfield units to stopping power ratios. The proton range in mice was determined and a 3D-printed mouse-like phantom was irradiated. In this thesis, determined stopping powers of all CIRS phantom inserts except for lung tissues were accurate with a maximum deviation of 2% or 6.5% after calibration with QuantumGx or VECTor micro-CT scanners. Larger deviations were observed for CIRS lung inserts or Gammex inserts. The irradiated gafchromic films inserted in the murinemorphic phantom showed dose distributions visualising the mouse’s anatomy, yet the dose was too low due to irradiation in the distal edge of the Bragg peak. The latter was confirmed with Monte Carlo simulations. This dosimetric set-up to place mice tumours in the proton Bragg peak should thus be slightly improved and can then be of great value for in vivo experiments of proton therapy. This enables execution of many new experiments with DNA damage repair inhibition and FLASH irradiations, potentially leading to an increase of the therapeutic window in proton therapy and less side effects for HNSCC-patients. ... Read more

Background and purpose: One of the main challenges in external beam radiotherapy treatment of locally advanced cervical cancer patients is dealing with bladder and rectum filling. Organ filling causes inter­fraction motion of the uterus, requiring large treatment planning volumes, or a plan library. Current assessment of tumor position is mainly done by visual inspection of a Cone Beam Computed Tomography (CBCT) scan. Eventually, this can lead to inter-­ and intra-­observer variability when choosing the best treatment plan from the plan library based on bladder filling. The incoming introduction of auto­contouring tools to obtain automatically-­generated (AG) contours of the bladder and the rectum on CBCT scans, allows the easier identification of these organs at risk and consequently, faster localization of the tumor region. However, to rely on these AG contours in the decision of plan selection, it is necessary to know if they have been reliably segmented. The goal of this project is to develop a strategy based on quantitative image features, to evaluate the quality of the AG contours to know if they are suitable for plan selection assessment.
Materials & Methods: 140 LACC patients from Erasmus MC were included. For each patient, bladder and rectum contours were obtained from each of the CBCT scans done throughout the treatment (five fractions (CBCT scans) per patient). These contours were automatically­generated using a deep learning­-based auto­segmentation algorithm. Gold­-standard contours were manually delineated in some CBCT scans, but the rest of the automatically­generated contours did not have the corresponding ground-­truth contour, hence they were labeled with a score between 1 (bad quality) and 5 (good quality). For consistency, gold­-standard contours were included in the dataset with the class label 5. The contours were relabeled to have a binary classification problem, and those with label 3 were removed. Each contour volume was divided into three subregions: core region, inner and outer shell. This contour data was used for a comparison study between two supervised machine learning (ML) methodologies:
Random forest (RF) networks and Logistic Regression (LR). For both strategies, feature extraction and selection were implemented. In RF methodology, a prior step of dimensionality reduction using principal component analysis (PCA) was performed. In LR, univariate feature selection followed by a multivariate logistic regression analysis was done. Before implementing the classifiers, the dataset
was split into a training set and a test set. The ML models were trained using the training set, and they were tested on new unseen data. Predictions on the test data were obtained and used for evaluation of the model's performance using evaluation metrics: accuracy, sensitivity, specificity, confusion matrix, ROC curve, and AUC.
Results: The RF classifier performed on the bladder test data with an AUC value of 0.87, while for the LR model, the value obtained was 0.77. The trained RF model identified the accurate and inaccurate bladder contours with a sensitivity of 94% and a specificity of 54%. The trained LR model resulted in a sensitivity of 91% and a specificity of 42%. In the case of the rectum, the RF classifier performance is indicated with the AUC value of 0.89, while the LR model obtained a value of 0.84. In the case of sensitivity and specificity, the RF model got 96% and 38%, and the LR classifier 95% and 38%, respectively.
Conclusion: Random forest classifiers give the best results in terms of performance and classification skills for the OARs considered, especially for the bladder. It has been demonstrated that quantitative image features, paired with the corresponding contour class label, can be used for deriving statistical relationships from the data. This allows the identification of contouring errors and classifying the contours based on their quality. With the increasing automation of different steps in the radiotherapy treatment workflows, the automatic contour QA tool developed would be a key step in the process to ensure a faster, more feasible, and consistent plan selection. The tool could act as a support tool for radiotherapy technicians when choosing the plan from the plan library that best fits the daily anatomy of the patient. ... Read more

Radiotherapy is among the most popular modalities used for cancer treatment. Proton therapy is a promising kind of radiotherapy, which uses protons characteristic maximum dose deposition to a specific tissue depth, for precise tumor irradiation. However, in comparison to conventional photon radiotherapy, proton therapy is sensitive to more treatment uncertainties, like errors in proton range and geometrical errors. To account for these uncertainties robust optimization and robustness evaluation have been developed, to obtain and guarantee the robustness of the treatment plans against potential error. Robust optimization uses a fixed number of scenarios, for which the plan is optimized using the worst-case scenario. However, a proper weighting of the sampled scenarios, with the corresponding probabilities is missing using this approach. Probabilistic treatment planning can resolve this limitation, but it requires significant time resources. To evaluate the quality of a treatment plan, one must quantify the effect of errors in dose distributions. However, dose distributions calculations are time-consuming, therefore in this work, polynomial chaos expansion (PCE) is used, as a dose meta-model, for fast and advanced dose analysis. This model uses a series of expansion in terms of polynomials, to evaluate the dose distribution when different errors occur. The first aim of this work was to improve the PCE construction speed and accuracy. To build the model a fixed number of dose scenarios from a dose engine are required. Currently, PCE dose meta-model is constructed using dose distributions with a 1% Monte Carlo (MC) noise level. The trade-off between the noise level (larger noise level allows a faster model) and the model's accuracy, was investigated. The accuracy for models built with larger noise levels (2 & 3%) was compared, to conclude that the default value is the most efficient choice. Additionally, PCE built using the dose differences (between a scenario with no errors and a shifted scenario), was used to improve models' accuracy for complex anatomies. However, probably due to the larger impact of MC noise when dose differences (smaller dose input value) are considered, the PCE accuracy was not high as expected. The second aim of this work is to evaluate robustness and trade-offs made in treatment planning in a clinically robust neuro-oncological patient, and a clinically complex patient case. As mentioned, robust optimization does not consider the occurrence probability of uncertainties. Therefore the treatment plans might be over-conservative. Using PCE robustness evaluations, we compare treatment plans with different robustness settings. From the results, we concluded that a further reduction of the settings was possible for the clinically robust patient. Third, for five robust skull base patients, we investigated the trade-off between homogeneity of target dose, robustness, and dose to healthy tissue. PCE robustness evaluations were used for comparison of inhomogeneous (120% maximum target dose) and homogeneous (107% maximum target dose) treatment plans, for different robustness settings. From these planning approaches, no intrinsic difference in the degree of robustness was observed. However, the inhomogenous treatment plan resulted in more healthy tissue sparing overall, at the expense of homogeneity of the target dose. The final aim of this project is to use PCE robustness evaluations towards a clinically feasible iterative probabilistic treatment planning. We attempted to find a linear relationship between the optimal robustness settings which meet a probabilistic goal, and a scaling factor, using iterations of PCE robustness evaluations. For our method, the overall scaling factor "α" was used, which is based on the assumptions that the actual margin recipe is approximately linear, and that the real underlying robustness recipe is a scaled version of the photon therapy margin recipe (van Herk's formula). The conclusion was that for one iteration an overall linear scale factor allows for a substantial gain. However, linear modeling and one iteration do not suffice to find the overall robustness settings optimum. ... Read more

Proton radiotherapy has a dosimetric advantage over photon therapy to spare healthy tissue closely positioned to the tumor mainly due to the absent exit dose. In The Netherlands, the Proton therapy centers currently take a relative biological effectiveness (RBE) of 1.1 compared to photons to deliver an iso-effective treatment. However, initial clinical evidence indicates a variable proton RBE in brain patients with the linear energy transfer (LET) as an important physical parameter. The LET significantly increases at the end of the radiation field, and contributes to an increased probability to develop brain lesions. With the introduction of radiation response models, the first goal of this thesis is to evaluate the impact of the RBE/LET effect in intensity-modulated proton therapy (IMPT) plans. Furthermore, the main goal is to reduce the RBE/LET effect in treatment planning.

We incorporated the probability of lesions origin (POLO) model published in literature to determine the RBE model-based normal tissue complication probability (NTCP) for three glioma patients treated with IMPT at HollandPTC, The Netherlands. The dose and LET distributions were computed using a Monte Carlo system. For the investigation of the RBE/LET effect in treatment planning, we modified several beam settings of the clinical IMPT plan, including the beam angle, beam energy, and robustness. Furthermore, we combined treatment modalities to reduce the NTCP.

We compared the results of the clinically used IMPT plan with the results obtained by the modified IMPT plans. The local redistribution of LETd leads to a decrease in NTCP up to the point when the LETd becomes uniform. The robustness did not reveal deviations in terms of the NTCP. By choosing appropriate beam angles that result in a smeared out
LETd distribution, the NTCP does not improve for small deep located tumors, improves relatively modest by 11.6% for elongated tumors, and significantly improves by 37.0% for large, superficially located tumors. The inclusion of partial transmission beams lowers the NTCP by 30-50% relative to the clinical IMPT plan while limiting the relative increase in mean brain dose by 5-16%. When comparing the IMPT plan with the photon plan used for plan comparison, the VMAT plan always results in the lowest NTCP and provides a relative improvement in NTCP by 60-75%. Meanwhile, the mean brain dose significantly increases by 50-80% compared to the clinical IMPT plan. Intermediate NTCP-Dmean(brain minus CTV) values are achieved when combining protons with the photons or by including proton transmission beams.

In general, we can conclude that the inclusion of partial proton transmission beams is more promising than choosing appropriate beam angles to lower the RBE/LETd effect. However, further optimization of transmission beams is required. Moreover, an improvement in NTCP is always at the cost of the mean dose to healthy tissue. On top, our results support further investigation to combine different modalities, like protons and photon fractionation. ... Read more