Author ORCID Identifier

https://orcid.org/0000-0002-0503-0175

Date of Graduation

5-2023

Document Type

Dissertation (PhD)

Program Affiliation

Biostatistics, Bioinformatics and Systems Biology

Degree Name

Doctor of Philosophy (PhD)

Advisor/Committee Chair

Clifton David Fuller

Committee Member

Wendy Woodward

Committee Member

Jason Stafford

Committee Member

Stephen Lai

Committee Member

Christine Peterson

Committee Member

Carlos Cardenas

Committee Member

David Fuentes

Abstract

Oropharyngeal cancer (OPC) is a widespread disease and one of the few domestic cancers that is rising in incidence. Radiographic images are crucial for assessment of OPC and aid in radiotherapy (RT) treatment. However, RT planning with conventional imaging approaches requires operator-dependent tumor segmentation, which is the primary source of treatment error. Further, OPC expresses differential tumor/node mid-RT response (rapid response) rates, resulting in significant differences between planned and delivered RT dose. Finally, clinical outcomes for OPC patients can also be variable, which warrants the investigation of prognostic models. Multiparametric MRI (mpMRI) techniques that incorporate simultaneous anatomical and functional information coupled to artificial intelligence (AI) approaches could improve clinical decision support for OPC by providing immediately actionable clinical rationale for adaptive RT planning. If tumors could be reproducibly segmented, rapid response could be classified, and prognosis could be reliably determined, overall patient outcomes would be optimized to improve the therapeutic index as a function of more risk-adapted RT volumes. Consequently, there is an unmet need for automated and reproducible imaging which can simultaneously segment tumors and provide predictive value for actionable RT adaptation. This dissertation primarily seeks to explore and optimize image processing, tumor segmentation, and patient outcomes in OPC through a combination of advanced imaging techniques and AI algorithms.

In the first specific aim of this dissertation, we develop and evaluate mpMRI pre-processing techniques for use in downstream segmentation, response prediction, and outcome prediction pipelines. Various MRI intensity standardization and registration approaches were systematically compared and benchmarked. Moreover, synthetic image algorithms were developed to decrease MRI scan time in an effort to optimize our AI pipelines. We demonstrated that proper intensity standardization and image registration can improve mpMRI quality for use in AI algorithms, and developed a novel method to decrease mpMRI acquisition time.

Subsequently, in the second specific aim of this dissertation, we investigated underlying questions regarding the implementation of RT-related auto-segmentation. Firstly, we quantified interobserver variability for an unprecedented large number of observers for various radiotherapy structures in several disease sites (with a particular emphasis on OPC) using a novel crowdsourcing platform. We then trained an AI algorithm on a series of extant matched mpMRI datasets to segment OPC primary tumors. Moreover, we validated and compared our best model's performance to clinical expert observers. We demonstrated that AI-based mpMRI OPC tumor auto-segmentation offers decreased variability and comparable accuracy to clinical experts, and certain mpMRI input channel combinations could further improve performance.

Finally, in the third specific aim of this dissertation, we predicted OPC primary tumor mid-therapy (rapid) treatment response and prognostic outcomes. Using co-registered pre-therapy and mid-therapy primary tumor manual segmentations of OPC patients, we generated and characterized treatment sensitive and treatment resistant pre-RT sub-volumes. These sub-volumes were used to train an AI algorithm to predict individual voxel-wise treatment resistance. Additionally, we developed an AI algorithm to predict OPC patient progression free survival using pre-therapy imaging from an international data science competition (ranking 1st place), and then translated these approaches to mpMRI data. We demonstrated AI models could be used to predict rapid response and prognostic outcomes using pre-therapy imaging, which could help guide treatment adaptation, though further work is needed.

In summary, the completion of these aims facilitates the development of an image-guided fully automated OPC clinical decision support tool. The resultant deliverables from this project will positively impact patients by enabling optimized therapeutic interventions in OPC. Future work should consider investigating additional imaging timepoints, imaging modalities, uncertainty quantification, perceptual and ethical considerations, and prospective studies for eventual clinical implementation. A dynamic version of this dissertation is publicly available and assigned a digital object identifier through Figshare (doi: 10.6084/m9.figshare.22141871).

Keywords

magnetic resonance imaging; artificial intelligence; deep learning; image processing; radiotherapy, oropharyngeal cancer

Share

COinS