Degree

Doctor of Philosophy (PhD)

Department

Physics & Astronomy

Document Type

Dissertation

Abstract

Anatomic changes occurring mid-treatment for patients undergoing radiation therapy for locally advanced lung cancer can degrade the quality of the intended treatment plan. These changes include tumor regression, geometric misalignment, and lung density changes (atelectasis and pleural effusion), and are visible on the daily 3D cone beam CT (CBCT). To maintain the intended treatment quality, adaptive radiotherapy (ART) can be employed to modify the treatment plan to account for these anatomic changes. However, the evaluation on when to adapt is currently completed manually by the treating clinicians, resulting in a subjective and inconsistent application. To address these limitations, a series of predictive ART models utilizing quantifiable metrics generated from daily CBCT imaging were developed to improve the efficiency and accuracy of identifying when ART should be triggered. These models were trained utilizing 1158 daily CBCT datasets from 43 patients who were treated for locally advanced lung cancer between 2010 – 2018. The “ground-truth” need for ART was systematically evaluated on each CBCT by two trained clinical radiation therapy staff, with ART indicated when anatomic deviations exceeded predefined tolerances across three consecutive days.The rate and magnitude of anatomic changes occurring during treatment were quantified using common image registration similarity criteria to calculate the anatomical differenced between the simulation CT and each on-treatment CBCT. These quantifying criteria, along with the “ground-truth” ART evaluation, were utilized as input data for predictive logistic regression models for lung ART, with one model for each of five types of observed anatomic change. The generalization of the predictive ART models was evaluated with a 10-fold cross-validation technique, focusing on the specificity, sensitivity, and first identified fraction of the model compared to the manually classified ground truth. Sensitivity (78.9% – 100%) and specificity (93.5% - 100%) was compared to published studies reporting the rate of manually identified patients, with three exhibiting significant improvement. Overall identification efficiency for the entire process added 52.4±17.7 seconds, significantly improving published rates ranging from 3 minutes – 7 minutes. Future work should include application of the models to a large independent dataset and investigating refinement of model predictions to direct dosimetric benefits for ART.

Date

8-24-2020

Committee Chair

Fontenot, Jonas

Available for download on Thursday, August 24, 2023

Share

COinS