SU‐E‐J‐208: Synthetic Datasets for Adaptive Radiotherapy Simulations in Lung Cancer Using Principal Component Analysis (PCA)

Purpose: To evaluate feasibility of PCA to produce realistic lung cancer datasets that demonstrate clinically relevant trends of tumor regression for simulating adaptive radiotherapy treatments. Methods: For four non small cell lung cancer patients, weekly helical CTs were acquired during the course of radiotherapy under active breathing control (ABC) and structures of interest were delineated by a qualified physician. Following a rigid alignment, week one images were registered to each of the remaining weeks using the ITK Demons algorithm. Resulting displacement vector fields (DVF) were reshaped into a composite matrix for PCA, with each row consisting of weekly displacements for homologous points. Intermediate and extrapolated PCA coefficients were drawn from a linear fit of the data and from a patient specific PDF constructed using kernel density estimation of the original coefficients. Principal components and the intermediate coefficients were employed to generate synthetic DVFs. A pseudo inverse DVF was generated from each intermediate synthetic DVF and used to warp week one images and contours for each patient resulting in eight sets of images and contours. Dice coefficients for week one risk structures and all propagated contours of the same are calculated in an effort to quantify the realistic nature of the images i.e. how well each represents a realistic pose of patient anatomy. Tumor volumes are calculated for each time point and are compared with clinically observed regression trends. Results: Dice coefficients for all (both linear and random samples) heart, left lung, and right lung contours were respectively: 0.9, 0.9, 0.9, for patient 1; 0.7, 0.6, 0.7, for patient 2; 0.6, 0.7, 0.7 for patient 3; and 0.3, 0.3, 0.2 for patient 4. Mean decrease in tumor volume was 36.2 percent. Conclusions: PCA has the potential to produce realistic poses of patient anatomy that exhibit clinically relevant trends of regression. Funded by P01‐CA116602.