Room: Exhibit Hall | Forum 7
Purpose: Functional imaging can identify the biology and characteristics of tumor cells. This facilitates personalized treatment planning to deliver escalated dose to these tumor pockets. In this study, we apply machine learning to uncover the relationship between tumor biology/characteristics and the tumor-control-probability(TCP) of an optimized treatment plan.
Methods: Three optimized-plans are compared: 1)Standard clinical plan, 2)2-stage TCP-driven PET-image-guided dose-escalated plan, and 3)Direct TCP-optimized PET-image-guided dose-escalated plan. The TCP formulation of these treatment models are derived from the Zaider and Minerbo models. A machine learning modeling framework is designed to uncover features of the tumor pockets and patient characteristics that can predict the TCP values. Five predictive models are used (â€œNeural Networkâ€?,â€œRegression SVMâ€?,â€œDecision Treeâ€?,â€œGaussian Process Regressionâ€?,â€œLinear Regressionâ€?) and eight patient-specific features (â€œPTV volumeâ€?, â€œPET volumeâ€?,â€œPET volume percentage in PTVâ€?,â€œShape factor of PETâ€?,â€œAverage distance from PET to bladder centerâ€?,â€œAverage distance from PET to rectum centerâ€?,â€œShortest distance from PET center to PTV boundaryâ€?, and â€œPET center to PTV center distanceâ€?) are constructed for 40 cervical cancer cases.
Results: Gaussian process regression using 4 features (â€œPTV volumeâ€?, â€œPET volume percentage in PTVâ€?, â€œshortest distance from PET center to PTV boundaryâ€? and â€œPET center to PTV center distanceâ€?) gives the best prediction. The mean square prediction error over all cases are 0.0007, 0.0025, 0.0000 respectively for the three treatment models. In particular, TCP prediction of direct TCP-optimized PET-image-guided dose-escalated plans always generates a consistently good accuracy, inferring that this treatment model is more robust and reliable and will result in a more reliable treatment outcome.
Conclusion: The model involves personalized tumor biology/physical/spatial characteristics. It allows quick evaluation of the (TCP-)quality of a treatment plan. The machine-learning modeling framework provides physicians a predictive platform to evaluate the potential TCP baseline of a patient. It also allows for efficient decision-making in selecting most effective (personalized) treatment plans.
Funding Support, Disclosures, and Conflict of Interest: This research is partially supported by a grant from the National Science Foundation.