PT  - JOURNAL ARTICLE
AU  - Eric Wolsztynski
AU  - Finbarr O'Sullivan
AU  - Janet O'Sullivan
AU  - Jingmei Jiang
AU  - Janet Eary
TI  - On the complementarity of structural and radiomic features for FDG-PET-based prognosis in sarcoma
DP  - 2018 May 01
TA  - Journal of Nuclear Medicine
PG  - 1730--1730
VI  - 59
IP  - supplement 1
4099  - http://jnm.snmjournals.org/content/59/supplement_1/1730.short
4100  - http://jnm.snmjournals.org/content/59/supplement_1/1730.full
SO  - J Nucl Med2018 May 01; 59
AB  - 1730Objectives: A methodology for spatial statistical modelling of the 3D FDG-PET tumor uptake distribution in sarcoma, developed recently by our group, yields a set of variables that capture structural aspects of this distribution. This set includes an assessment of structural heterogeneity and of local metabolic gradients. Here we explore the potential complementarity of this quantitation strategy with established texture features commonly used in radiomics. Prognostic potential of multivariate combinations of both sets of features is further assessed via various statistical schemes, including a number of machine learning models. Methods: We explore a quantitative analysis of a cohort of 198 primary sarcoma studies imaged with FDG prior to treatment at the University of Washington. The correlation structure of this quantitation, which includes both methodologies alongside routine patient and clinical variables (age, gender, metabolic volume, SUVmax, etc.), is explored in terms of correlation structure and factor analysis, in order to evaluate and explain their potential complementarity. Outputs from a number of feature selection techniques based on both classical statistical and machine learning strategies are also compared. These techniques include stepwise selection from Cox and logistic models, simulated annealing, support vector machines, random forests, neural networks and other schemes. The corresponding output multivariate models are assessed for prognostic utility in terms of Cox proportional hazard models, in order to identify clinically relevant feature sets. Results: Exploratory analysis of this multivariate set indicates that structural and textural features span different areas of the quantitative information space. This finding suggests that each of these two PET-based quantitative methodologies captures different aspects of the intratumoral FDG uptake distribution. This work also demonstrates that feature selection for multivariate prognostic assessment varies with the selection strategy. However, in a large majority of scenarios, cross-validated feature selections tend to associate both structural and textural features for improved multivariate prognostic assessment. Incidently, this study also demonstrates that radiomic features derived from FDG-PET imaging have prognostic utility in sarcoma in a multivariate setting that does not contain structural the novel descriptors. Conclusion: Statistical assessment of the volumetric structure of the FDG uptake distribution within primary sarcoma tumors is relevant for both tumor characterization and prognosis. Our findings suggest that it complements typical radiomic summaries based on texture features for characterization of tumor metabolism, and that this complementarity contributes to improved prognostic modelling from the baseline scan. Research Support: Supported in part by Science Foundation Ireland grant PI 11/1027 and by the National Cancer Institute grant ROI-CA65537