tesis de maestría
Impact of dose and grading uncertainties on xerostomia prediction using machine learning classification
Fecha
2020Registro en:
10.7764/tesisUC/FIS/49526
Autor
Hormazábal González, María Constanza
Institución
Resumen
Due to the sharp dose gradients present in IMRT treatments, small uncertainties can generate significant differences between the planned and delivered dose distribution for head-and-neck cancer patients. This thesis investigated the impact of dosimetric and toxicity grading uncertainties on the prediction of post-radiotherapy xerostomia development.
Dosimetric uncertainties were simulated by Gaussian shifts of the planned dose for a cohort of 77 patients. We analyzed demographic, radiomic and dosimetric features as predictor of xerostomia. Univariate and multivariate studies were carried out and compared with classic logistic regression prediction models. These models were based on the nominal features extracted from the planned dose and the mean and standard deviation of features extracted from the shifted dose. In the case of grading uncertainties, we introduced a sample weight in the predictive model.
The predictive power of the models was quantified in terms of AUC. In general, no change in the AUC values was observed when considering the mean of the features. Nevertheless, when considering the standard deviation, we obtained models with higher AUC, especially in the classical model analysis, corresponding to models based on mean doses. The model based on ipsilateral and contralateral mean doses improved its performance from an AUC of 0.47 (0.44-0.50) to 0.83 (0.80-0.85). In the univariate analysis we found that the dose gradient in the patient's right-left direction was highly correlated with xerostomia development. However, this predictor becomes a bad indicator of xerostomia when considering the dose uncertainties. In the multivariate study, developing machine learning models with good performance was possible, reaching AUC values close to 0.9.
The uncertainties in xerostomia grading showed to influence the probability space generated by the predictive model. However, these results were not significant, obtained the same AUC values as those calculated without the grading uncertainties.