Integration of artificial intelligence in quality assurance for head and neck cancer radiotherapy clinical trials Du Wang

Format:

Book

Thesis/Dissertation

Author/Creator:

Wang, Du, author.

Contributor:

University of Pennsylvania. Bioengineering., degree granting institution.

Language:

English

Subjects (All):

Bioengineering.

Biomedical engineering.

Therapy.

0202.

0541.

0212.

0800.

Local Subjects:

Bioengineering.

Biomedical engineering.

Therapy.

0202.

0541.

0212.

0800.

Genre:

Academic theses

Physical Description:

1 online resource (162 pages)

Contained In:

Dissertations Abstracts International 87-12A

Place of Publication:

Ann Arbor : ProQuest Dissertations and Theses, 2026

Language Note:

English

Summary:

Deviations in radiotherapy (RT) delivery in multi-institutional head and neck cancer (HNC) clinical trials can compromise treatment quality and clinical outcomes, yet the quality assurance (QA) processes designed to detect such deviations remain labor-intensive, subjective, and fragmented across stages of the workflow. This dissertation investigated whether artificial intelligence (AI) can be systematically integrated into the QA workflow of NRG Oncology HNC clinical trials to provide automated, objective, and connected assessment spanning contour review, treatment plan evaluation, and outcome analysis.Three interconnected aims were pursued. Aim 1 evaluated auto-segmentation models (Carina, Therapanacea, and in-house nnU-Net/nnFormer) as independent computational references for contour QA. Using 187 NRG-HN005 cases to establish Score 1-based thresholds and 173 RTOG 0522 cases for cross-trial validation, the framework achieved high sensitivity for detecting Score 3 deviations in well-defined OARs (e.g., BrainStem: 100% sensitivity, 10.9% FPR; Cochlea: 100% sensitivity, 13.2% FPR; Thyroid: 83.3% sensitivity, 9.1% FPR), supporting the use of AI as a first-pass screening tool for contour review. Aim 2 evaluated whether a knowledge-based planning (KBP) model trained on curated NRG-HN001 plans could provide anatomy-specific benchmarks for treatment plan QA. In the NRG-HN001 validation cohort, 66% of KBP-reoptimized plans improved sparing of at least seven OARs, with parotid mean dose reductions of 2.2-3.8 Gy and optic structure reductions of 5.7-6.6 Gy. Cross-trial application to 50 RTOG 0522 cases revealed statistically significant dose reduction potential across all seven evaluated OARs, with the largest improvements for parotid glands (Dmean -7.9 to -9.3 Gy), larynx (Dmean -8.8 Gy), and BrainStem (D0.03cc -10.9 Gy, all p < 0.001). Aim 3 developed interpretable single-task XGBoost and multi-task gradient boosting models to predict 14 binarized clinical outcomes using clinical, QA, and dosimetric features from 565 RTOG 0522 patients. The best overall prediction was achieved for 2-year overall survival (AUC = 0.781). Expanded dosimetric feature sets derived from AI-generated contours improved prediction for selected toxicity endpoints, including mucositis (AUC 0.597 vs 0.459, BH p = 0.018) and hypoacusis, while manually reviewed original contours remained advantageous for some survival endpoints.Collectively, these findings support a multi-layered AI-augmented QA framework in which contour assessment, plan evaluation, and outcome modeling inform one another across the RT trial workflow. Rather than replacing expert review, this framework can improve QA consistency, identify clinically meaningful deviations, and guide future protocol design

Notes:

Source: Dissertations Abstracts International, Volume: 87-12, Section: A.

Advisors: Xiao, Ying Committee members: Zhu, Timothy C.; Teo, Boon-Keng Kevin; Mcbeth, Rafe; Avery, Stephen M.

Ph.D. University of Pennsylvania 2026

Vendor supplied data

Local Notes:

School code: 0175

ISBN:

9798247983361

Access Restriction:

Restricted for use by site license