Vuong D, et al. Med Phys 2020.
BACKGROUND: Radiomics is a promising tool for the identification of new prognostic biomarkers. Radiomic features can be affected by different scanning protocols, often present in retrospective and prospective clinical data. We compared a CT based radiomics model based on a large but highly heterogeneous multi-centric image dataset with robust feature pre-selection to a model based on a smaller but standardized image dataset without pre-selection.
MATERIALS AND METHODS: Primary tumor radiomics was extracted from pre-treatment CTs of IIIA/N2/IIIB NSCLC patients from a prospective Swiss multi-centric randomized trial (npatient = 124, ninstitution = 14, SAKK-16/00) and a validation dataset (npatient = 31, ninstitution = 1). Four robustness studies investigating inter-observer delineation variation, motion, convolution kernel and contrast were conducted to identify robust features using an intra-class correlation coefficient threshold > 0.9. Two 12-months overall survival (OS) logistic regression models were trained: (1) on the entire multi-centric heterogeneous dataset but with robust feature pre-selection (MCR) and (2) on a smaller standardized subset using all features (STD). Both models were validated on the validation dataset acquired with similar reconstruction parameters as the STD dataset. The model performances were compared using the DeLong test.
RESULTS: In total, 113 stable features were identified (nshape = 8, nintensity = 0, ntexture = 7, nwavelet = 98). The convolution kernel had the strongest influence on the feature robustness (less than 20% stable features). The final models of MCR and STD consisted of one and two features, respectively. The features of STD were identified both as non-robust. MCR did not show performance significantly different from STD on the validation cohort (AUC [95%CI] = 0.72 [0.48-0.95] and 0.79 [0.63-0.95], p=0.59).
CONCLUSION: Prognostic OS CT radiomics model for NSCLC based on a heterogeneous multi-centric imaging dataset with robust feature pre-selection performed equally well as a model on a standardized dataset.