From Dose Engine to Machine Learning in Radiation Therapy: Frameworks, Insights, and Clinical Applications

Abstract

This dissertation consists of two main parts. The first part, spanning Chapters 2–5, represents the primary research project conducted during the author’s Ph.D. study and focuses on the development of a pencil-beam–based dose calculation engine and machine-learning (ML)–inspired optimization frameworks for IMRT and VMAT. Collectively, this work is analogous to building a mini-in-house treatment planning system (TPS) from scratch.Chapter 2 presents the development of a fast analytical dose calculation engine. The fitting of analytical pencil-beam kernels and open-field fluence intensities using a 10-control-point model is described. An analytical ray-tracing–based correction is implemented to account for density heterogeneity. Transformations among key coordinate systems used in DICOM and treatment planning systems are formulated. Chapter 3 introduces an ML-inspired IMRT optimization framework that reformulates fluence map optimization (FMO) as the training of a single-layer network. It enables direct use of modern ML toolkits such as PyTorch, with optimization performed using built-in L-BFGS optimizer. Results demonstrate that it achieves plan quality comparable to classical gradient-descent–based methods and superior performance in certain cases. It also exhibits faster convergence, improved robustness to randomized initializations, and the ability to explore uneven beam-weight distributions that are difficult for classical methods to identify, making it a practical and effective alternative to conventional approaches. In Chapter 4, the ML-inspired framework is extended to VMAT optimization by reformulating direct aperture optimization (DAO) as a multi-layer neural network training. Optimization parameters, including multileaf collimator (MLC) leaf positions and control-point weights, are encoded into parameterized activation layers and a final weighting layer, respectively. In addition to hard constraints such as jaw boundaries and leaf-pair collision avoidance, machine delivery limits including maximum dose rate and gantry speed are incorporated as regularization terms and optimized jointly with the clinical objectives. Using the L-BFGS optimizer in PyTorch, the optimization converges successfully. Resulting plans are converted to DICOM, imported into Eclipse without errors, and delivered on a TrueBeam without interlocks. Dose comparisons with benchmark IMRT plans are consistent with clinical experience, demonstrating the feasibility of applying ML-inspired methodologies to VMAT optimization. Chapter 5 presents an extended study enabled by the developed dose engine and optimization framework. The search space of IMRT FMO is analyzed using both theoretical considerations and numerical perturbation analysis. Results reveal that the optimization landscape consists of multiple flat plateaus at different levels where classical first-order methods are prone to premature convergence due to vanishing gradients. Second-order methods that account for curvature can converge more efficiently. It provides new insights into IMRT optimization behavior and is made possible by the fast-converging optimization platform developed in this work. Chapters 6 and 7 constitute the second part of the dissertation and address two clinically aligned research topics stemming from the primary project. Dose rate and gantry speed are key variables for VMAT delivery; however, for other treatment modalities such as IMRT and total body irradiation (TBI), dose rate is restricted to a limited set of discrete presets, reducing delivery flexibility. Chapter 6 investigates an MLC-modulation–based method for dose-rate modulation on linear accelerators. In this approach, MLC leaves are programmed to traverse distances exceeding their physical speed limits, prompting the machine to automatically reduce the dose rate to maintain accurate monitor-unit delivery. The algorithm is successfully demonstrated for IMRT, 3D-CRT, and TBI fields across dose rates ranging from 1 MU/min to 600 MU/min. Increased delivery variance is observed and quantified for dose rates below 50 MU/min. Inspired by the dose engine developed in the first part of the dissertation, Chapter 7 presents an accuracy study of photon-counting CT (PCCT)–based dose calculation. Three approaches are evaluated: a ground-truth method based on electron density (Rho), a measured Hounsfield unit lookup table (HLUT), and a rapid alternative using Eclipse’s default HLUT for energy-integrated CT. In-silico dose calculations are performed for pancreatic cancer cases using 70-keV virtual monoenergetic images (VMI). Results show that PCCT dose calculation using the measured HLUT closely matches the ground truth, while the rapid alternative also demonstrates good agreement, as confirmed by 3D gamma analysis. Electron-density images effectively remove contrast-agent effects, suggesting that dose calculation on Rho images with visualization on VMI is a recommended strategy for treatment planning in the presence of contrast agents. Chapter 8 presents the vision for future directions and potential extensions of this project. It outlines several avenues for continued development, including components that are currently under active investigation as well as ideas that remain conceptual and may be pursued in future work by the author or other researchers. Together, these perspectives aim to position the framework as an evolving research platform with long-term potential for methodological, computational, and clinical expansion.