ESTRO 2025 - Abstract Book

S3714

Physics - Radiomics, functional and biological imaging and outcome prediction

ESTRO 2025

Keywords: Radiomics, Image Biomarker, PET SUV

References: 1. A. Zwanenburg et al., The Image Biomarker Standardization Initiative: Standardized Quantitative Radiomics for High-Throughput Image-based Phenotyping, Radiology 295, 328–338 (2020). 2. P. Whybra et al., The Image Biomarker Standardization Initiative: Standardized Convolutional Filters for Reproducible Radiomics and Enhanced Clinical Insights, Radiology 310, e231319 (2024). 3. Z-Rad, https://github.com/medical-physics-usz/z-rad

277

Digital Poster Development of Multi-organ Dual-Omics machine learning models for predicting trismus in post radiotherapy nasopharyngeal carcinoma patients Si Wing Tsui 1,2 , Jiang Zhang 2 , James C. H. Chow 1 , Karus P. S. Yeung 3 , Haylie P. Y. Wong 1 , Winkie W. K. Chong 1 , Tiffany K. K. Tse 3 , Alex K. C. Leung 1 , Ka Man Cheung 1 , Kwok Hung Au 1 , Benny C. Y. Zee 4 , Wai Tong Ng 5 , Wing Cheung Vincent Wu 2 , Jing Cai 2 1 Department of Clinical Oncology, Queen Elizabeth Hospital, Hong Kong, Hong Kong. 2 Department of Health Technology and Informatics, The Hong Kong Polytechnic University, Hong Kong, Hong Kong. 3 Department of Occupational Therapy, Queen Elizabeth Hospital, Hong Kong, Hong Kong. 4 Centre for Clinical Research and Biostatistics, Jockey Club School of Public Health and Primary Care, The Chinese University of Hong Kong, Hong Kong, Hong Kong. 5 Department of Clinical Oncology, School of Clinical Medicine, The University of Hong Kong, Hong Kong, Hong Kong Purpose/Objective: Radiation-induced trismus can be a serious but common late complication in post-radiotherapy (RT) nasopharyngeal carcinoma (NPC) patients (1). Its relationship with masticatory muscles and RT dose lacks investigation in clinical research (2). Assessing the anatomical characteristics of masticatory organs-at-risk (OARs) and impact of radiation doses for pre-treatment identification of high-risk patients are of significant interest. This study aimed to develop radiomics and dosiomics signatures of multiple masticatory OARs and construct machine learning models to predict late RT-induced trismus in NPC patients. Material/Methods: A total of 86 NPC patients who were treated with intensity-modulated RT (≥66Gy) and disease-free for at least 3 years were retrospectively included. Clinical information, CT and planning dose maps were collected. Six masticatory structures were contoured: temporo-mandibular joint, mandible, lateral pterygoid, medial pterygoid, messeter and temporalis muscles. Dental gap was measured according to Orastretch Jaw MIO Scale by clinician, and a value of <35mm was considered as positive in trismus (3). A comprehensive set of radiomics features were extracted from CT images and dose maps within each OAR. Independent relevant features were selected using outcome ANOVA test and feature clustering. Signatures were developed for each data modality and OAR by logistic regression modeling on patient clustering results and further selected based on outcome relevancy. Four machine learning models (logistic regression, random forest, gradient boosting, SVM) were developed using clinical variables and significant radiomics or/and dosiomics signatures. Model performance metrics, including area under the receiver operating characteristic (ROC) curve (AUC), classification accuracy (CA), F1 score, precision and recall value were evaluated under 100-time repeated 3-fold cross-validation. Results: Of 86 patients, 20 patients (23.3%) were positive in trismus. Univariate logistic regression showed that none of the clinical features (n=16) were statistically correlated with this complication. A total of 14 radiomics and dosiomics signatures were developed, and 12 were found as significant with the clinical outcome. Among all the four machine