TH-AB-304-12: Validation of a Morphological Xerostomia Prediction Model

Authors


Abstract

Purpose:

We present a validation study of a morphological NTCP model predicting dry mouth syndrome (xerostomia) based on the spatial distribution of the dose received by salivary the glands in head-and-neck cancer patients during radiation therapy.

Methods:

We examined a recently proposed model based on Bayesian multivariate logistic regression. It uses scale-invariant moments of the three-dimensional dose distribution to explicitly include information about the distribution of the dose within the parotid glands in different spatial directions. We tested the classification performance of this model on our independent patient cohort and compared it to routinely used clinical mean dose models. The predictive power of the models was evaluated using receiver operating characteristics (ROC).

Results:

The study comprised 105 head-and-neck cancer patients from varying tumor sites. Of these patients, 13 reported severe xerostomia between five and 48 months after radiation therapy. The AUC of the morphological model, the mean-dose model based on the contralateral parotid gland, and the mean-dose model based on both parotid glands were 0.79, 0.71, and 0.69 respectively. The morphological model performed better than the model based on the mean-dose in the contralateral gland (p-value=0.04) as well as the model based on the mean-dose in both parotid glands (p-value=0.05). Preliminary results regarding the underlying feature selection reconfirm the high predictive power of the second moment, i.e. the spread of the dose, in cranio-caudal direction within the contralateral parotid gland.

Conclusion:

Our study confirms the benefit of spatial dose information over volumetric dose information for the prediction of xerostomia after radiation therapy. Incorporation of such morphological models during treatment planning could improve the evidence for clinical decision making and patients’ quality of life after treatment. In the future, we will explore the possibility to additionally incorporate parotid shrinkage, tumor regression, uncertainties stemming from imperfect delineations, patient positioning and organ motion.

Ancillary