Room: Karl Dean Ballroom C
Purpose: This study proposes and validates a machine-learning approach utilizing patient anatomical changes to predict parotid mean dose changes, thereby facilitating plan adaptation for H&N cancer.
Methods: Parotid mean dose changes are assumed to correlate with patient anatomical changes, quantified by 65 features in three sets: SET1) body shrinkage, resulting in beam path changes to the parotid; SET2) parotid movement with respect to the high-dose region; SET3) parotid volume changes. SET1 is quantified using features measuring the signed-distance between the parotid and skin. SET2 quantifies parotid movement using the signed-distance of the parotid to two PTV position surrogates: C2 dens and basilar tip of the occipital bone. A decision-tree classifier is built from a database of plans from 18 patients (36 parotids) previously treated with adaptive radiotherapy. Inputs are a patientâ€™s features calculated from the initial planning CT and secondary image set (CT/CBCT/MRI). Output is â€œ1â€? if the predicted parotid mean dose from the secondary image set is x% larger than its original value. During feature selection, leave-one-out cross validation combined with enumerating k-combinations (k=1, 2, 3 and 4) of the 65 features is used to find a feature subset maximizing classifierâ€™s accuracy. Gini-diversity index and maximal number of decision splits=10 are used in the classifier.
Results: When predicting x=5% (or x=10% as shown in the below parentheses) mean dose increase, two SET2 features (one SET1 feature and one SET2 feature) yielded maximal accuracy of 86.1% (86.1%) with 30.5% (19.4%) over prevalence=55.6% (66.7%); TPR=87.5% (75%), TNR=85% (91.7%), PPV=82.3% (81.8%) and NPV=89.5% (88%).
Conclusion: Using specific quantitative features characterizing patient anatomical changes, the classifier effectively predicts x=5% and x=10% parotid mean dose increase with an accuracy of 86.1%.