Predicting Plan Quality From Patient Geometry: Feature Selection and Inference Modeling
D Ruan*, W Shao, J DeMarco, P Kupelian, D Low, UCLA Department of Radiation Oncology, Los Angeles, CASU-E-T-214 Sunday 3:00:00 PM - 6:00:00 PM Room: Exhibit Hall
Purpose: To investigate and develop methods to infer treatment plan quality from the geometric features of PTV/OAR structures; to discover and identify features of high prognostic values.
Methods: This study explores the prognostic utility of geometric features of two categories: (1) absolute geometry, characterizing the volumes of single structures (PTV, OARs); and (2) relative geometry, based on the minimal 3D distance and/or overlapping volume between pairs of structures. Using prostate as a pilot site, we developed inference models to 'predict' SBRT plan quality of DVH end points. We developed and assessed (1) a full linear regression model based on both absolute and relative geometric features, (2) a sparsity-penalized linear regression model, (3) a linear regression model based on absolute geometry features only; (4) a learning-based nonparametric model. Cross-validation was used for both selecting the parameter values as well as quantifying the inference performance. The best inference method for each of the DVH end points was identified to reveal the structural and prognostic differences among them.
Results: For linear regression, using sparsity-regularization discovered geometric features that were mostly absolute, demonstrating their dominant linear prognostic utility. However, introducing relative geometric features improved the plan quality prediction by 15% for all DVH end points. In contrast, nonparametric models had a heavier dependence on relative geometry features. While linear regression based on both features sets predicted OAR DVH points slightly better, the nonparametric method excelled in predicting PTV coverage and conformality.
Conclusions: The inference result from this study provides an 'expectation' for the plan quality before the planning is to be performed, providing reference goals for the planner and a baseline for detecting abnormality. The use of relative geometry complements the absolute geometry with information on spatial configuration of the PTV/OAR structures of individual patients, and the variation in achievable conformality as a consequence.