160 - Subregional Parotid Dosiomics, DVH Integration and Prediction of Xerostomia after Radiotherapy for Nasopharyngeal Carcinoma

08:10am - 08:20am ET

Room 156

Presenter(s)

Hongfei Sun, PhD - Xijing Hospital, Xian 710032, Shaanxi

H. Sun¹, Z. An², L. N. Zhao³, and J. Zang⁴; ¹Department of Radiation Oncology, Xijing Hospital, Air Force Medical University, xi'an, Shaanxi, China, ²Department of Radiation Oncology, Xijing Hospital, Fourth Military Medical University, Xi'an, Shaanxi, China, ³Department of radiation oncology, Xijing Hospital, the Fourth Military Medical University, Xi'an, China, ⁴Department of Radiation Oncology, Xijing Hospital, Air Force Medical University, Xi'an, China

Purpose/Objective(s):

To investigate whether integrating dosiomics features and conventional dose-volume histogram (DVH) metrics derived from parotid gland subregions improves prediction of xerostomia after radiotherapy for nasopharyngeal carcinoma (NPC), and to compare predictive performance between whole-gland and subregional models.

Materials/Methods:

A total of 400 NPC patients treated with intensity-modulated radiotherapy were retrospectively analyzed, including 168 patients with xerostomia and 232 without xerostomia after =12 months of follow-up. The parotid gland was anatomically divided into three subregions: superficial lobe, superior lobe, and stem-cell–enriched region. From the whole gland and each subregion, dosiomics features were extracted from 3D dose distributions, together with conventional DVH parameters. Following feature selection, dosiomics and DVH features were integrated for model development. Five machine learning algorithms—Logistic Regression (LR), Extra Trees (ET), Random Forest (RF), Support Vector Machine (SVM), and XGBoost (XGB)—were trained and evaluated using 5-fold cross-validation. Model discrimination was assessed using area under the receiver operating characteristic curve (AUC).

Results:

Across 5-fold cross-validation, the lowest mean validation AUC was observed in the whole-parotid model (0.652, ET), whereas the highest mean AUC was achieved in the superior-lobe model (0.705, LR). The superficial lobe and stem-cell region yielded mean AUCs of 0.692 (ET) and 0.676 (RF), respectively. The best single-fold performance was observed in the superficial-lobe model (AUC = 0.765, ET), while the worst performance was observed in the whole-parotid model (AUC = 0.549, SVM). Subregional combined models consistently outperformed whole-gland combined models.

Conclusion: Integrating dosiomics with DVH metrics enables effective prediction of xerostomia in NPC patients. Subregional modeling of the parotid gland provides superior discriminative performance compared with whole-gland analysis, supporting the presence of spatial heterogeneity in parotid radiation response. External validation is needed to confirm the robustness of this subregional dosimetric strategy.