Ai-Enabled Online Plan Adaptation For Mr-Guided Stereotactic Ablative Radiotherapy (Sabr) of Head and Neck Cancer

<p>Head and neck cancer (HNC) is a prevalent cancer type worldwide. Stereotactic ablative radiotherapy (SABR) has emerged as an effective treatment for HNC, delivering highly conformal doses to the tumor target while sparing surrounding normal tissues with a sharp dose gradient. However, the accuracy of the treatment delivery is limited by setup errors, anatomical changes, and intra-/inter-fraction organ motion. The emergence of MR-guided adaptive radiotherapy (ART) has the potential to further improve the SABR of HNC, by providing superior visualization of soft tissue and enabling real-time plan adaptation based on the daily anatomical changes of patients. This novel technology has the potential to achieve improved local control and normal tissue sparing in HNC patients. However, its clinical implementation is challenged by the available technologies of accurate delineation of gross tumor volume (GTV) for treatment planning and the real-time accurate contouring for MR-based adaptive planning. The purpose of this study is to address the challenges associated with MR-guided online ART for HNC SABR by automating the key steps in the clinical workflow using deep learning methods.</p> <p>First, we developed an advanced auto-segmentation framework to automate the GTV delineation for SABR treatment planning. The framework was specifically designed to simulate the GTV delineation process performed by radiation oncologists, and multimodality images (CT, PET, and MRI) were included to improve the contouring accuracy. We found that more than 95% of the automatically generated GTVs were clinically acceptable.</p> <p>Next, we automated the deformable image registration between planning and daily images, enabling accurate and rapid contour propagation for the MR-guided online ART. A novel hierarchical registration framework was proposed and validated for CT-to-MR and MR-to-MR deformable image registration. Our evaluation demonstrated superior performance than traditional registration tool implemented in current clinical practice.</p> <p>The last component of this project was to automate the generation of high-quality synthetic CT from daily MR images for MR-based adaptive planning and dose calculation. To achieve this, we developed a novel deep learning model based on Cycle Generative Adversarial Network. We validated both image quality and dosimetric accuracy of the generated synthetic CT images by comparing them to their corresponding real CT images.</p> <p>In conclusion, we have developed and validated the advanced deep learning methods for GTV autosegmentation, deformable image registration, and synthetic CT generation. These tools enabled automation of key steps in the MR-guided online ART workflow for HNC SABR, improving the accuracy and efficiency of treatment planning and delivery.</p>

Name thesis:degree_name

Doctor of Philosophy (PhD)

Level thesis:degree_level

Dissertation (PhD)

Year dc:date.available

2023

Authors dc:creator

• Zhao, Yao
• <p>0000-0001-8863-841X</p>

Contributors dc:contributor

• Jinzhong Yang
• Laurence Court
• Xin Wang

• Deep learning
• MR-guided radiotherapy
• Stereotactic ablative radiotherapy
• Head and neck cancer
• adaptive radiation therapy
• Health and Medical Physics

Repository record dc:identifier

https://digitalcommons.library.tmc.edu/utgsbs_dissertations/1283

OAI identifier oai:identifier

oai:digitalcommons.library.tmc.edu:utgsbs_dissertations-2340