308. Evaluation of the stability and discriminative power of PET radiomic features in phantom studies

Abstract

Purpose PET radiomics is a promising and emerging field with large potential for tumor phenotype characterization. Radiomic Metrics (RM) analysis, nonetheless, is not yet standardized and the impact of the whole processing chain on RM hast not yet been clearly understood. In this work we characterized the impact of the following major issues on RM: (i) four different discretization methods, (ii) acquisition duration, (iii) reconstruction algorithm and iv) image filtering. The influence of the region volume and Standardized-Uptake-Value (SUV) on the RM was also measured. Methods and materials The CGITA software v1.4 was used to estimate 42 RM, from an uniform phantom and a purpose built “heterogeneous” one. Statistical effect size indicators were used to determine for each RM: (i) the impact of each nuisance factor, and (ii) the discriminating power. Results The discretization procedure performed at the beginning of RM quantification is critical with respect to the dependency of RM values on lesion volume and SUV. The method based on relative range had overall better proprieties than those based on bins of fixed width in units of SUV. Based on the robustness with respect to all factors, we categorized RM based on the influence of the imaging workflow and on the ability to better discriminate heterogeneous PET patterns. The families of RM less robust resulted to be the Gray Level Run Length Matrix (GLRLM) and the Neighborhood gray tone difference matrix (NGTD) ones while the Gray Level Co-occurrence Matrix (GLCM) group as well as the SUV first order RM were the most suitable. Conclusion From the experiments performed in this work, processing pipeline recommendations were provided, together with a list of 20 sufficiently robust and discriminative RMs for PET radiomic analysis. This work was supported by AIRC IG18965.