Improving relative stopping power calibration in X-ray treatment planning CT using sparse proton radiographies

Butz, Ines (2026): Improving relative stopping power calibration in X-ray treatment planning CT using sparse proton radiographies. Dissertation, LMU München: Faculty of Physics

Abstract

Radiation therapy with heavy charged particles is an emerging modality of cancer treatment, offering the potential to deliver highly conformal dose distributions to the irradiated tumor while sparing surrounding healthy tissue. To fully exploit this advantage over conventional photon beam therapy, precise knowledge of the ion beam range within the patient is essential. To this end, the ability of certain irradiated tissue to stop ions, i.e., the tissue stopping power, needs to be determined to a high degree of accuracy. Currently, treatment planning is based on an X-ray computed tomography (CT) of the patient acquired prior to treatment, which is converted to stopping power relative to water based on a scanner-specific semi-empirical calibration curve. This conversion procedure introduces uncertainties into treatment planning due to associated measurement uncertainties but also due patient-specific deviations from the generic conversion curve. Ion transmission imaging, as the native imaging modality for particle therapy, can directly probe the patient-specific stopping power relative to water (RSP) distribution. While tomographic ion imaging is technically challenging, RSP information obtained from ion radiography acquired from few projection angles can be fused with the known patient anatomy from the X-ray CT. This can be achieved by generating virtual radiographies, simulated from the X-ray CT relying on forward projection models and an initial estimate of the conversion curve, and comparing these to the measured radiographies. The conversion curve can then be adapted to minimize the difference between measured and simulated radiographies. This optimization procedure is, however, challenged by the forward projection model being based on simplifying assumptions, and therefore not perfectly describing the full physics underlying ion radiographic imaging. This work aims to improve the RSP calibration of the treatment planning CT using sparse proton radiographies and consists of three in-silico studies based on Monte Carlo (MC)-simulated radiographic data. In the first part, potential enhancements to the analytical forward operator are investigated, employing a differentiable implementation in cases where the forward operator is formulated as a function of RSP, either inherently or due to the removal of simplifying assumptions. Additionally, improvements to the data representation of the measured radiographies are considered taking into account different detector types. Instead of compressing the radiographic measurement information into two-dimensional (2D) projection images, optimization can be performed using a higher-dimensional, detector-specific data representation, yielding higher calibration accuracy. Using radiographies acquired with idealized single particle tracking detectors, RSP accuracy in the order of 0.2% is achieved, where modifications to the forward operator merely yielded small further improvements at high computational cost. Optimizing the RSP conversion directly on the detector signal obtained with an idealized particle integrating detector combined with a differentiable analytical forward model achieved remarkably low calibration errors in the order of 0.5%, albeit still at high computational cost. Currently available multi-layer ionization chambers provide comparable effective depth sampling suggesting that similar calibration accuracy may be achievable with real detectors. Secondly, a deep learning-based forward model is trained with the aim to replace the analytical operator and improve as well as speed up calibration optimization. Powerful sequence models are investigated for the task of modeling accurate detector dose distributions based on the traversed patient anatomy. While results achieved with a transformer-based architecture are promising, further work is needed to achieve the targeted accuracy of the predicted dose. Interfacing the trained forward model with a conventional iterative optimization workflow demonstrates practical feasibility of the envisioned approach, with advantages in terms of computational effort specifically for detectors providing 2D or three-dimensional (3D) dose measurements. In a third part, a data-driven end-to-end approach is developed, predicting the calibrated CT image directly from an initial guess and the acquired radiographic data without the need for iterative optimization. Even though the approach is based on radiographic data in form of projection images, it outperforms analytical optimization based on higher-dimensional structured radiographic data for two out of the three considered idealized detector configurations. Furthermore, the data-driven approach offers the advantage of significantly reduced computation times, which will be an important factor for clinical application of proton radiography as in-room image guidance modality. This work advances the state of the art in conventional proton radiography-based iterative RSP calibration optimization by demonstrating the advantages of advanced forward projection models in combination with higher-dimensional, structured projection data. Furthermore, it introduces a novel data-driven RSP calibration approach that combines proton radiography-derived RSP information with the treatment planning CT through deep learning. The present in silico study relies on MC-simulated radiographies generated under idealized detector assumptions in a line-scan imaging setup. Future work may extend these approaches to 2D imaging scenarios as well as to ex vivo and in vivo proton radiographies.

• Item Type: Theses (Dissertation, LMU Munich)
• Keywords: proton therapy, relative stopping power calibration, proton radiography, deep learning
• Subjects: 500 Natural sciences and mathematics; 500 Natural sciences and mathematics > 530 Physics
• Faculties: Faculty of Physics
• Language: English
• Date of oral examination: 4. May 2026
• 1. Referee: Parodi, Katia
• Signature of the printed copy: 0001/UMC 32011