Knowledge modeling for computer aided treatment planning

Abstract

The purpose for this study is to develop robust and comprehensive knowledge models which can provide patient specific prediction of achievable dose distribution in radiation therapy plans for a wide range of cancer types. These models are useful tools to guide radiation therapy planning. Clinical RT plans for a number of cancer types including prostate, head and neck, anorectal, lung and spinal SBRT plans are studied retrospectively. The knowledge modeling correlates patient anatomical features with the dose features embedded in the RT plans. A number of patient’s anatomical and dosimetric features are considered in the model. The geometrical relationships between the OARs and PTV are represented by the distance-to-target histogram, DTH and the distance- to-OAR histogram. Important anatomical and dosimetric features were extracted from DTH and DVH by principal component analysis (PCA). For spine SBRT plans, voxel-level dose distribution is also predicted in addition to the dose-volume histogram (DVH) of spinal cord in order to explore the tradeoff between PTV coverage and spinal cord sparing by an active optical flow model (AOFM). A step-wise multiple regression method was used to select the most significant patient features which influence the dose distribution. To validate the knowledge models, the model predicted dosimetric parameters in the OARs of the validation cases are compared with the actual plan values. The difference between the model predictions and the actual values for some example dosimetric indexes are (mean±s.d.: volume in percent of OAR volume, dose in percent of prescription dose): lung V5Gy in lung IMRT plans: 0.1±6.9, Parotid median dose in head and neck IMRT plans: 0.9±5.6, Spinal cord D2% in spinal SBRT plans: 0.1±0.6. The predicted dosimetric indexes by the knowledge models can provide good estimates for those in actual plans. These models can help to improve the quality and efficiency of treatment planning.