PET beyond pictures

Abstract

PET imaging data analysis provides crucial instruments for diagnosis and therapeutic assessment of many forms of cancer. The utility of simple statistical summaries of the metabolic volume and distribution of image intensities for both baseline prognosis and therapeutic response assessment is now well established and prompted the definition of PET-specific guidelines (PERCIST) for routine practice. Current methodologies however fail to exploit the complexity of such data effectively. A number of quantitative methodologies have been developed in recent years that capture different aspects of the PET uptake information. Converging efforts in the radiology community have led to the emergence of radiomics, which consists of high-throughput extraction of data-driven pseudo-markers and model-building based on intensive machine learning techniques, in order to build more effective prognostic models. In parallel, other forms of statistical analysis of the structural characteristics of the distribution of PET uptake are also developed, with the aim to provide complementary description in clinical terms. Both types of approaches can be considered and combined to form a more informed, patient-specific evaluation of risk and therapeutic assessment. In this chapter we describe these quantitation methodologies and illustrate them on clinical examples. We provide a high-level review of major methodological aspects of quantitative PET-based tumor characterization, highlight key components of the imaging data acquisition and post-processing chain that may influence such analyses, and discuss the role of artificial intelligence and machine learning frameworks in the context of sarcoma and melanoma.