161 - Evaluating CT-based Radiomics and Dosiomics Outcome Models for Locoregionally Advanced Head and Neck Cancer: A Retrospective Analysis of the NRG/RTOG 0522 trial

08:20am - 08:30am ET

Room 156

Presenter(s)

Xin Tie, PhD - University of Pennsylvania, Philadelphia, PA

X. Tie¹, D. Wang², S. H. Lee², H. Geng³, H. Zhong⁴, K. Men⁵, D. I. Rosenthal⁶, J. J. Caudell⁷, A. K. Singh⁸, C. U. Jones⁹, S. Rudra¹⁰, S. Brule¹¹, T. J. Galloway¹², G. Shenouda¹³, P. R. Anne¹⁴, C. Cardenas¹⁵, A. S. Garden⁶, R. M. Lanning¹⁶, Q. T. Le¹⁷, and Y. Xiao²; ¹University of Pennsylvania, Philadelphia, PA, ²Department of Radiation Oncology, Perelman School of Medicine, University of Pennsylvania, Philadelphia, PA, ³University of Pennsylvania, Perlman School of Medicine, Philadelphia, PA, ⁴Department of Radiation Oncology, University of Pennsylvania, Philadelphia, PA, ⁵State Key Laboratory of Molecular Oncology and Department of Radiation Oncology, National Cancer Center/National Clinical Research Center for Cancer/Cancer Hospital, Chinese Academy of Medical Sciences (CAMS) and Peking Union Medical College (PUMC), Beijing, China, ⁶Department of Radiation Oncology, The University of Texas MD Anderson Cancer Center, Houston, TX, ⁷Department of Radiation Oncology, H. Lee Moffitt Cancer Center and Research Institute, Tampa, FL, ⁸Department of Radiation Medicine, Roswell Park Comprehensive Cancer Center, Buffalo, NY, ⁹Sutter Medical Center Sacramento, Roseville, CA, ¹⁰Department of Radiation Oncology, Winship Cancer Institute of Emory University, Atlanta, GA, ¹¹The Ottawa Hospital Cancer Program, Ottawa, ON, Canada, ¹²Fox Chase Cancer Center, Philadelphia, PA, ¹³McGill University Health Center, Montreal, QC, Canada, ¹⁴Thomas Jefferson University, Philadelphia, PA, ¹⁵University of Alabama at Birmingham, Birmingham, AL, ¹⁶Department of Radiation Oncology, University of Colorado School of Medicine, Aurora, CO, ¹⁷Stanford University, Stanford, CA

Purpose/Objective(s):

Radiomics and dosiomics have shown promise for outcome prediction in locoregionally advanced head and neck cancer (LA-HNC). However, previous work has typically relied on single-institution datasets, raising concerns about feature robustness. This study aims to analyze whether dose-volume histogram (DVH) features, CT-based radiomics and dosiomics offered added prognostic value beyond clinical variables in predicting overall survival (OS), locoregional failure-free survival (LFFS), and distant metastasis-free survival (DMFS).

Materials/Methods:

The retrospective analysis included 669 patients receiving chemo-radiotherapy from the multicenter RTOG 0522 trial. Clinical variables included demographics (age, sex), treatment arm and modality, common risk factors (tumor stage, tumor site and size, smoking status, Zubrod status, hemoglobin level), and prescription dose. DVH features were extracted from the high-dose planning target volume (HDPTV), elective-dose PTV (EDPTV), and four protocol-specified organs at risk (spinal cord, parotid glands, larynx, brachial plexus). Radiomics and dosiomics were extracted for HDPTV and EDPTV (including intensity, shape, and texture) from planning CT images and 3D dose maps, leading to a total of 400 features. A standardized outcome modeling pipeline was developed, comprising feature selection and penalized Cox regression with elastic net regularization (CoxNet). Models were trained and evaluated using fivefold nested cross-validation, repeated 20 times to reduce variability from data partitioning. Performance was assessed by concordance index (C-index), and differences between models were evaluated by Mann-Whitney U tests. Kaplan-Meier (KM) analyses assessed risk stratification based on the median predicted risk from the training folds, with log-rank tests evaluating separation.

Results:

CoxNet models built on clinical and DVH features achieved the highest performance for OS (C-index 0.71±0.01), and DMFS (0.62±0.01), significantly outperforming clinical-only models (0.68±0.01 for OS; 0.57±0.01 for DMFS; both P<0.001). Adding radiomics and dosiomics did not improve OS (0.71±0.01) or DMFS (0.60±0.01) prediction. For LFFS, the model incorporating all features had the highest C-index (0.59±0.01) but did not significantly surpass the clinical-only or clinical+DVH models (both 0.58±0.01). KM analyses showed significant separation between low- and high-risk groups stratified by the clinical+DVH models (log-rank P<0.001 for OS and DMFS; P=0.011 for LFFS).

Conclusion:

In this large multicenter clinical trial cohort, DVH features provided added prognostic value beyond clinical variables for OS and DMFS, whereas radiomics and dosiomics did not offer further benefits. These findings suggest that standard dosimetric parameters are more robust for risk assessment in LA-HNC. Future studies will focus on prospective validation and the integration of multi-omics data to improve LFFS prediction.