Generative AI for cone-beam CT dose reduction and intensity correction in adaptive radiotherapy

Chan, Yan Chi (2025): Generative AI for cone-beam CT dose reduction and intensity correction in adaptive radiotherapy. Dissertation, LMU München: Medizinische Fakultät

Abstract

In image-guided radiotherapy (IGRT), cone-beam computed tomography (CBCT) is used to align patients in the treatment position. CBCT scans administer radiation exposure and subject patients to secondary cancer risk. However, lowering CBCT imaging dose continues to be challenging as the image quality degrades. In current practice, the frequency of scanning can be limited, leading to a potential decrease in localisation precision. Moreover, in cases of adaptation, patients are required to undergo an addition planning CT due to insufficient CBCT image quality, resulting in extra radiation exposure and prolonged treatment time. This thesis reports on investigations towards the minimum CBCT imaging dose without loss of accuracy in terms of synthetic CT (sCT) image generation using generative artificial intelligence (AI) for adaptive radiotherapy (ART). Many studies translated full dose CBCT images into sCT images using deep learning (DL) algorithms such as U-Net, cycle-consistent generative adversarial network (cycleGAN) or contrastive unpaired translation (CUT). However, only few studies investigated the potential of low imaging dose CBCT. In this thesis, the lowest achievable CBCT imaging dose for online adaptation was investi- gated. Compared to the previous studies, this work provides a structured investigation on imaging dose levels (100%, 25%, 15%, 10%) and evaluations with IGRT-related metrics, including patient positioning, treatment dose calculations and organ contouring. Online adaptation in IGRT is currently limited by the CBCT image quality. In the first part of this thesis, the basics of cancer and IGRT in chapter 1, the adaptation workflow in chapter 2 and the physics of CBCT in chapter 3 are discussed, respectively. This will serve as an introduction to explain how CBCT-guided IGRT works and to identify the challenges of using CBCT in ART. While DL is used to enhance full dose CBCT images, generating sCT images from low dose CBCT requires additional under-sampling streaks removal. The improvements and recent studies for DL-enabled full dose CBCT-to-CT translation are discussed in chapter 4. In chapter 5, the significance of low imaging dose CBCT, and the synthesis of low imaging dose sCT images and the DL algorithms that can be used are discussed. In chapter 6, the metrics for evaluating sCT images are discussed. To investigate the minimum CBCT imaging dose for IGRT adaptation, we conducted two studies with generative AI models using a retrospective prostate patient dataset. In chapter 7, the patient database and the contributions of the two studies are explained. In chapter 8, the published papers for each of the studies for reference are attached. Especially in the second study, it was found that 25% is the minimum CBCT imaging dose for accurate treatment dose calculation and organ contouring when using the AI methods selected in this project. Finally, in chapter 9 Discussion, the findings and limitations in this work, the challenges that hinder the development of low imaging dose CBCT, and possible future works that can extend this study and facilitate clinical implementations of the new low imaging dose CBCT technique in the ART workflow are discussed.

• Dokumententyp: Dissertationen (Dissertation, LMU München)
• Keywords: CBCT, deep learning, imaging dose, adaptive radiotherapy
• Themengebiete: 600 Technik, Medizin, angewandte Wissenschaften ; 600 Technik, Medizin, angewandte Wissenschaften > 610 Medizin und Gesundheit
• Fakultäten: Medizinische Fakultät
• Sprache der Hochschulschrift: Englisch
• Datum der mündlichen Prüfung: 17. Oktober 2025
• 1. Bericht­erstatter:in: Landry, Guillaume