Abstract
Purpose: To introduce and validate a model of systematic tissue displacement based upon population statistics for prostate cancer patients undergoing fractionated external‐beam radiotherapy. Methods: Our dataset consisted of 19 patients, each with 9–14 pelvic fan‐beam CTs taken throughout treatment, all manually contoured and aligned to bone. Each patient's fractional images were deformably registered to his planning image using intensity and contour matching. The resulting displacement vector fields (DVFs) were averaged to find the patient‐specific tissue systematic error, which maps the patient's mean treatment anatomy to his planning anatomy. Then, each patient's prostate, rectum, and bladder were mapped from his planning image to its counterpart on an average reference image set using contour‐driven deformable registration. These planning‐to‐reference DVFs were used to transport the 19 systematic DVFs (limited to the space occupied by the organs) into the common reference frame. A principal component analysis (PCA) was performed to determine the principal modes of systematic DVF variability. A leave‐one‐out approach was used to test the hypothesis that systematic errors for patients not included in the training set can be accurately represented as a linear combination of PCA eigenvectors. The known average anatomy and the PCA reconstructed anatomy was compared using the Dice similarity coefficient. Results: The PCA showed that 92% of systematic error variations can be accounted for using 7 eigenmodes. Using the leave‐one‐out approach to reconstruct the patient's average anatomy, the mean (standard deviation) Dice for the prostate, bladder, and rectum were 0.91±0.03, 0.93±0.03, and 0.84±0.05 respectively. Conclusion: The PCA population modeling of systematic errors can accurately represent systematic error DVFs for arbitrary patients. Our model can be used to create an ensemble of randomly sampled systematic pelvic organ DVFs for use in probabilistic planning when the patient's actual systematic error is unknown a priori. Supported by NIH Grant P01 CA 116602; Supported by NIH Grant P01 CA 116602
