DOI
https://doi.org/10.25772/ZDHZ-KD97
Defense Date
2022
Document Type
Dissertation
Degree Name
Doctor of Philosophy
Department
Medical Physics
First Advisor
Christopher L. Guy
Abstract
Head and neck cancer is commonly treated with a six- to seven-week course of radiotherapy, during which a patient’s anatomy may change substantially, due to target reduction or weight loss. Anatomical changes lead to reduction in treatment quality due to decreased setup reproducibility and altered dose deposition compared to the original plan. Few clinics have developed a standard method for triggering resimulation and replan due to anatomic changes. This work investigates a new method for determining when to resimulate and replan HNC patients by utilizing their topographic anatomical changes to predict differences in planned versus delivered dose distributions. The first part of the work presents a method for deformable image registration of CT to CBCT which addresses the challenges of inaccurate Hounsfield units and truncated field of view present in CBCT. The registration method was validated on 10 HN patients using contour comparison, with average DSC of 0.82, 0.74, 0.72, and 0.69 for mandible, cord, and left and right parotid. The registration method was then used to generate dose maps and surface contours for 47 patients for the second part of this work, the development of a U-Net which takes the original dose distribution, the original surface, and the treatment day surface as input and predicts the treatment day dose distribution as output. The average RMSE and MAE between the true and predicted dose distributions for a test set of 6 patients was 4.25 and 2.15. This work proves feasibility of a dose prediction neural network using surface imaging.
Rights
© The Author
Is Part Of
VCU University Archives
Is Part Of
VCU Theses and Dissertations
Date of Submission
12-15-2022