3620 - Parotid Voxel Dose Analysis with Machine Learning Toward Xerostomia Prediction in Head and Neck Cancer Irradiation

M. Chao¹, J. Wei², T. Liu¹, and J. A. Penagaricano³; ¹Icahn School of Medicine at Mount Sinai, Department of Radiation Oncology, New York, NY, ²City College of New York, New York, NY, ³Department of Radiation Oncology, H. Lee Moffitt Cancer Center and Research Institute, Tampa, FL

Purpose/Objective(s): Xerostomia induced by the irradiation of salivary glands significantly impacts a patients quality of life and remains a clinical challenge despite technological advances in radiotherapy. The current practice by the mean dose constraints within the salivary glands neglects the local dose effect thus its efficacy is severely limited. This study aims to perform a voxel-based analysis of the spatial dose within the parotid gland to investigate the local dose patterns associated with xerostomia and predict its occurrence with machine learning methods. Materials/

Methods: A cohort of 240 patients who underwent intensity-modulated radiation therapy treatment were retrospectively selected for this study. A binary classification was adopted for predictive modeling: group 1 (119 patients) if grade 0 or 1 and group 2 (121 patients) if grade 2 or 3. The anatomy of the parotid gland from the computed tomography (CT) of these patients was standardized to that of a reference patient with rigid and multistage B-spline deformable registration. The resultant deformation field vectors were subsequently utilized to warp the parotid doses to the reference ones. To facilitate pattern recognition, the bilateral parotid doses were further regrouped into contra- and ipsi-lateral depending on their proximity to the tumor targets. Four highly competitive supervised machine learning (ML) models including ridge regression (RR), random forest (RF), Extra Trees (ET), and XGBoost (XB), together with ReliefF for feature importance determination, were employed to determine the feature importance of each voxel and discover the dose patterns within the glands that are correlated with xerostomia. The model performance was assessed by four scores: the area under the curves (AUC) of the receiver operating characteristic, the prediction, recall, precision, and F1 score.
Results: The four ML models were assessed with 5-fold cross-validation showing comparable performance, the AUC scores were 0.620 ± 0.099, 0.710 ± 0.089, 0.642 ± 0.092, 0.672 ± 0.084 for RR, RF, ET, and XB, respectively, and the averaged F1 score was around 0.6. The importance patterns showed that the doses to the anterior-lateral region of the contralateral parotid gland were observed to be more correlated with xerostomia from the multiple classifiers in addition to ReliefF.
Conclusion: The voxel-based analysis incorporating the spatial information of dose within the parotid gland has demonstrated an improvement in xerostomia predictive modeling as compared to the mean dose predictor. It is useful to study the local dose effect on xerostomia. With more patient data and independent studies, the findings from this investigation could be incorporated into treatment planning of head and neck patients to reduce the toxicity occurrence.