3750 - Advancing Image-Based Assessment of Long-Term Parotid-Gland Toxicity in Head and Neck Cancer Radiotherapy: Ultrasound Radiomics and Machine Learning

J. Wang¹, M. Chao¹, J. Zhang¹, Y. Lei², J. T. Liu³, J. W. Shelton², D. S. Yu², X. Yang², and T. Liu¹; ¹Icahn School of Medicine at Mount Sinai, Department of Radiation Oncology, New York, NY, ²Department of Radiation Oncology, Winship Cancer Institute of Emory University, Atlanta, GA, ³Department of Radiation Oncology, Icahn School of Medicine at Mount Sinai, New York, NY

Purpose/Objective(s): To introduce an ultrasound-based method for evaluating long-term radiation-induced parotid gland toxicity in head and neck cancer (HNC) patients. The method integrates ultrasound radiomics analysis of the parotid gland with machine learning techniques. Materials/

Methods: A cohort of 56 patients who had previously undergone radiotherapy for HNC was enrolled during routine follow-up visits (median: 25 months, range: 12 -87 months) when each participant underwent a sagittal ultrasound scan of the parotid glands, using a conventional scanner equipped with a linear probe (10 MHz). The scan duration is less than 5 minutes. Manual contours of the parotid glands were delineated on the B-mode images, and subsequently, radiomic features, were extracted. These features encompassed first-order histogram, second-order gray level and gray tone features. Patient self-assessment questionnaires rated saliva consistency and mouth dryness on a 0-3 scale. Correlations between extracted radiomic features, mean radiation dose, and toxicity outcomes were analyzed. Six machine learning models classified toxicity into none/mild versus moderate/severe, evaluated via area under the curve of the receiver operating characteristic curve (AUC-ROC), accuracy, precision, and sensitivity with three-fold cross-validation.
Results: Significant correlations (p < 0.05) were found between 34 radiomic features and mean dose (6.4 – 61.8 Gy). Notable features included Difference Variance of the gray-level co-occurrence matrix (Pearson correlation coefficient: 0.54, p < 0.01). For saliva consistency toxicity, 53 significant features were identified, with Root Mean Squared (Spearman correlation coefficient: 0.50, p < 0.01) being prominent. Mouth dryness showed 18 significant features, with Large Area Emphasis being notable (Spearman correlation coefficient: 0.30, p = 0.01). Table 1 presents the performance of the top machine learning models in classifying none/mild versus moderate/severe toxicities.
Conclusion: Ultrasound radiomics combined with machine learning effectively assesses long-term radiation-induced parotid toxicity in HNC radiotherapy, highlighting clear associations between radiomic features, radiation dose, and toxicity outcomes. This imaging method offers potential for improved patient care through enhanced monitoring and management of radiation-induced parotid toxicity. Abstract 3750 – Table 1: Top 2 machine learning models to classify none/mild vs. moderate/severe parotid toxicities

Saliva Consistency
Model	AUC-ROC	Accuracy	Precision	Sensitivity
Random Forest	0.77 ± 0.03	0.71 ± 0.03	0.76 ± 0.08	0.64 ± 0.05
KNN	0.70 ± 0.11	0.72 ± 0.10	0.71 ± 0.07	0.78 ± 0.17
Mouth Dryness
Model	AUC-ROC	Accuracy	Precision	Sensitivity
KNN	0.69 ± 0.14	0.63 ± 0.17	0.64 ± 0.19	0.58 ± 0.17
Decision Tree	0.53 ± 0.13	0.55 ± 0.09	0.57 ± 0.09	0.61 ± 0.09