Survival analysis in radiomics: a statistical learning approach on 18F-FDG-PET/CT imaging data of patients affected by multiple myeloma

Medical imaging data provide a valuable insight into oncological diseases by unlocking richer information for all phases of tumour management. Imaging is fundamental for tumour diagnosis and staging, precise localisation of tumoural lesions, treatment planning and monitoring, and follow-up. When dealing with tumours, it is of utmost importance to identify the Region Of Interest (ROI): image segmentation techniques play a key role in order to select the tumoural region to be analysed. Once the segmentation is obtained, radiomics aims at extracting quantitative features, able to comprehensively describe the ROI in terms of shape, voxel intensities, and texture. Multiple myeloma is a haematological cancer targeting plasma cells and forming lesions mainly in bone marrow, but also in paraskeletal areas and sometimes in extra-medullary areas. 18F-FDG-PET/CT imaging technique can detect lesions in different areas, distinguish between active and inactive myeloma, and assess the tumour metabolic activity, providing a comprehensive overview of the disease and of the response to therapy. Being an aggressive disorder, with no available cure, the disease is under study within the cutting-edge European GenoMed4All project. The aim of this work was to retrieve survival predictions by a Cox model starting from PET/CT imaging data alone. The clinical endpoint of interest was the Progression-Free Survival. The first steps involved image processing operations; a key role is played by the segmentation of the ROI, which must be anatomically meaningful and accurate, from the input PET/CT scans. In case of patients affected by multiple myeloma, the ROI was identified in the skeleton with a particular focus on the spine. Then, radiomic features were extracted and used as input covariates for a penalised Cox model to predict survival outcomes. The model performance, evaluated by concordance index, resulted in high metric values and was robust with respect to the cross-validation strategy.