Author ORCID Identifier
https://orcid.org/0000-0003-1414-3849
Date of Graduation
8-2018
Document Type
Dissertation (PhD)
Program Affiliation
Medical Physics
Degree Name
Doctor of Philosophy (PhD)
Advisor/Committee Chair
Laurence E. Court
Committee Member
Michalis Aristophanous
Committee Member
Jinzhong Yang
Committee Member
Arvind Rao
Committee Member
Clifton D. Fuller
Abstract
Head and neck intensity modulate radiation therapy allows for the delivery of high-precision radiotherapy by conforming radiation dose to the defined treatment targets achieving more accurate target dose distribution and better sparing of normal tissues. However, producing very precise treatment plans may be ineffective if the target volumes are not defined accurately. Furthermore, there are several reports of significant inter-observer variability when delineating these target volumes for head and neck cancers making this variability one of the largest sources of uncertainty in head and neck radiation therapy.
The purpose of this study was to develop algorithms to automate target delineation for oropharyngeal cancer patients. Automating this delineation process could aid in reducing inter-observer variability and provide a venue for head and neck target delineation standardization in radiation therapy. These algorithms would be especially valuable for head and neck cancers where the observed variability is highest amongst radiation oncologists.
An assessment of our head and neck section’s inter-observer clinical target volume delineation variability was conducted to quantify the variability in our algorithm’s inputs. We then developed two novel deep learning algorithms to auto-delineate high-risk and low-risk clinical target volumes. The predicted delineations for high-risk and low-risk clinical target volumes performed well in comparison to their respective ground-truth delineations. The quantitative analysis showed that the predicted volumes provided, on average, improved delineations when compared to the assessed inter-observer variability. Lastly, we investigated dosimetric differences on target coverage and normal tissues based on the physician delineated and deep learning auto-delineated low-risk target volumes. The percent volume receiving 95% of the prescribed dose on the original physician PTVs was found acceptable, per RTOG 1016 guidelines, on over 70% of auto-delineated plans. In addition, we found no significant difference in normal tissue doses between the physician and auto-delineated target plans.
This study resulted in strong evidence that auto-delineated clinical target volumes could aid in the standardization of target delineation in radiation therapy. The target volume auto-delineation algorithms showed an improvement in overlap and dosimetric agreement with respect to the reported variability in the literature. Future studies may validate the clinical use of these algorithms.
Keywords
deep learning, head and neck cancer, delineation, clinical target volumes