Use of Generative Disease Models for Analysis and Selection of Radiomic Features in PET

Radiomic positron emission tomography (PET) image features are increasingly used in conjunction with machine learning to predict clinical disease measures. However, a thorough understanding of these image features remains challenging due to their relatively high complexity, hampering a-priori selection of optimal features and model parameters for a predictive task. In this paper, we explore the use of a generative disease model (GDM) for feature analysis. The GDM generates a series of synthetic PET images that simulate progressive disease-induced changes in radiotracer binding. These images can be used to obtain the expected values of image features, estimate the effect of various parameters on the feature correlation with clinical measures, and to select optimal features prior to testing them on real data. As an illustrative example, we apply the GDM-based approach to brain PET imaging of Parkinson's disease subjects. Following initial validation, we use the GDM to understand the trends of change in the measured feature values with disease progression. Interestingly, the GDM revealed many features to change nonmonotonically, even with monotonic changes in radiotracer binding. An important implication of this finding is that different features may be optimal as biomarkers at different disease stages.