TITLE:
Predicting Parotid Dose Changes in Head-and-Neck Radiotherapy Using Machine Learning: Leveraging Anatomical Variations
AUTHORS:
Binbin Wu, Peng Zhang, Pengpeng Zhang, Gig Mageras, Jillian Tsai, James Mechalakos, Margie Hunt
KEYWORDS:
VMAT, Parotid, Predict, Machine Learning, Adaptive
JOURNAL NAME:
International Journal of Medical Physics, Clinical Engineering and Radiation Oncology,
Vol.13 No.4,
November
19,
2024
ABSTRACT: Purpose: This study describes a machine-learning approach utilizing patients' anatomical changes to predict parotid mean dose changes in fractionated radiotherapy for head-and-neck cancer, thereby facilitating plan adaptation decisions. Methods: Parotid mean dose changes during treatment sessions are assumed to correlate with patients’ anatomical changes, quantified by 65 geometrical features in four sets. SET1 is the parotid volumetric changes; SET2 is the distance changes from the parotid to the PTV; SET3 is the length of beam path changes between the parotid and skin near the neck; SET4 is the distance changes from the parotid to the two bony landmarks—the dens of the C2 and tip of the basilar part of the occipital bone. The introduced landmarks in SET4 are used as surrogates for the PTV in SET2 due to PTV’s unavailability at the simulation stage. Signed Euclidean distance is applied to quantify the distance and beam path length. A decision tree classifier to predict an x% increase in parotid mean dose is developed. In a study involving 18 patients (36 parotids) previously treated with adaptive radiotherapy, a leave-one-out cross-validation combined with enumerating 4 combinations of the 65 geometrical features is used to find a feature subset maximizing classifier’s accuracy. The classifier’s accuracy, with and without SET2’s PTV features inclusion, is evaluated to determine the SET4’s bony landmark surrogate feasibility. Results: Under x = 5% (or x = 10%) parotid mean dose increase: without SET2’s PTV features inclusion, one beam path feature from SET3 and one bony landmark feature from SET4 yield maximal accuracy of 86.1%, which is a 30.5% (19.4%) improvement over prevalence = 55.6% (66.7%); TPR = 87.5% (75%), TNR = 85% (91.7%), PPV = 82.3% (81.8%) and NPV = 89.5% (88%). With SET2’s PTV features inclusion, accuracy increases from 86.1% to 91.6%. Conclusion: Under the current 18 enrolled patients’ data, we found that the introduced SET4’s bony landmarks are feasible surrogates for the SET2’s PTV features in determining the parotid’s position variations relevant to the high dose region. The contours of the parotid, skin near the neck, and bony landmarks in initial/adaptive images are required to predict parotid dose changes. Further study with a larger sample size of patients is necessary to provide a robust basis for generalization.