New prognostic metastatic phenotypes based on the analysis of whole-body PET images using Artificial Intelligence

Cancer deaths occur in the vast majority of cases in patients with metastatic disease. Currently, although the prognosis depends on the number and type of organs affected, on the degree of invasion and the survival can range from weeks to years, patients with metastatic disease are all grouped into Stage IV of the American Joint Committee on Cancer (AJCC). In addition to the molecular profile of the disease (when available) that currently orients the therapeutic choice, knowing the patient's prognosis more accurately would allow physicians to optimize patient’s management. To better map the different metastatic profiles, the NEMO-PET project aims to identify new prognostic phenotypes for metastatic patients based on a comprehensive analysis of whole-body PET/CT images. The objective is to exploit the potential of machine learning methods to extract prognostic representations from signals measured in pathological and non-pathological regions and from their associations. First, in addition to clinical and biological data, features automatically extracted from the images will be aggregated in order to identify distinct metastatic phenotypes (clusters) and to establish their relationship with the prognosis of patients. Then, we will integrate features measured in non-tumor regions over the whole-body (eg, metabolism of lymphoid organs, sarcopenia), in order to identify their prognostic role, or even to highlight interactions that could refine the metastatic phenotype. We will demonstrate the relevance of these approaches for the characterization of metastatic breast and lung cancers in a bicentric context.