(Wednesday, 8/1/2018) 7:30 AM - 9:30 AM
Room: Karl Dean Ballroom C
Purpose: Prediction of spatially variant response to radiation therapy (RT) on longitudinal imaging can inform risk-adaptive treatment strategies, including biological target dose escalation and functional tissue avoidance. We developed a modeling framework for predicting spatially variant tumor response on early-RT FDG-PET/CT imaging while accounting for auto-correlation of neighboring voxels.
Methods: Nineteen patients with stage IIB-IIIB non-small cell lung cancer enrolled on the FLARE-RT trial underwent FDG-PET/CT imaging prior to (preRT-PET) and during week 3 (midRT-PET) of concurrent chemoradiotherapy. PreRT-PET and midRT-PET voxel grids were co-registered and resampled to the planned RTdose voxel grids within metablic tumor volumes (MTV) to ensure isomorphic voxel mapping. Generalized least squares (GLS) regression was utilized to predict MTV-voxel midRT-SUV from patient-level (tumor etiology, MTV preRT-SUV�ᵉᵃ�) and voxel-level (MTV-voxel preRT-SUV) predictors. To account for distance-dependent correlation of neighboring voxels and GLS residuals, Matérn covariance matrices were fit to patient-specific empirical variograms. Variogram-weighted regression coefficients and variogram-corrected standard errors were estimated using GLS. Mean absolute prediction error (MAE) of each GLS model was calculated using leave-one-patient-out cross-validation.
Results: Over 19 patients and 8516 voxels, MTV-voxel SUV decreased from 6.3 g/mL to 3.6 g/mL at midRT-PET on average, with large inter-patient (standard deviation [SD]: 2.9 g/mL) and intra-tumor variability (SD: 1.7 g/mL). GLS voxel residuals were spatially correlated (R>0.1) within a model-estimated distance of 4 cm. Patient-level GLS prediction produced MTV-voxel midRT-SUV MAE of 2.5 g/mL while combined patient- and voxel-level GLS prediction achieved lower MAE of 1.9 g/mL (Wilcoxon sign-rank p<0.001). MTV-voxel preRT-SUV was the most important individual predictor of midRT-SUV.
Conclusion: Variogram-weighted generalized least squares regression provides a statistical framework to predict voxel-level response patterns during therapy. The multiscale model can be extended to predict post-treatment tumor response and normal tissue functional changes, and forms the basis to inversely optimize spatial-temporal dose distributions for precision radiotherapy.
Funding Support, Disclosures, and Conflict of Interest: This work was supported by NIH/NCI R01CA204301.
Keywords
PET, Statistical Analysis, Lung
Taxonomy
IM- PET : Machine learning, computer vision
Contact Email
