Radiomics feature analysis for survival prediction in multiple myeloma: An automated PET/CT approach

Background and objective: 18F FDG PET/CT imaging is emerging as a tool for Multiple Myeloma (MM) evaluation. The goal of this study is to assess the most informative regions and features in MM patients using a fully automatic imaging-based pipeline by evaluating the performance of survival models with radiomics features for risk stratification. Methods: A dataset of whole-body-18F FDG PET/CT images from 227 patients was used. All images were first automatically segmented to find the spine and the remaining skeletal structures. Several masks were then generated using this segmentation, and radiomic features were extracted from both images applying these masks. The features were aggregated to create three datasets per mask: one per imaging type and a dataset combining both. Four survival model families were trained on these datasets to estimate progression-free survival. In total, 128 model × image-origin × mask configurations were tuned to maximize mean c-index via five-fold cross-validation. The best performing model that utilized both imaging modalities was then analyzed to assess the prognostic power of its features. Results: To achieve the best ranking performance, the best configuration was a DeepSurv model trained solely with features from the PET image extracted using a mask created by using the spine and the region around and inside it. It achieved a mean c-index of 0.657 and IBS of 0.197 using 5-fold cross-validation. Image-texture features were the most relevant for risk prediction, with higher heterogeneity correlating with increased risk of an event. The pipeline is available in the project's GitLab repository. Conclusions: The best performing models were random survival forests and DeepSurv, which outperformed Cox-based linear models. Models with both CT and PET provided better results than with just one, on average, however the most informative imaging technique was PET, with more features present when combining both images and the overall best performing model. When building the masks, the inclusion of the paramedullary region was important, although adding the whole skeleton did not improve results. The most relevant features for survival analysis were textural features, with image heterogeneity correlating with a higher risk of disease progression.