Progression-Free Survival Prediction in Head and Neck Cancer using Fused PET-CT Radiomics and Machine Learning

It is essential to have a pipeline for predicting the survival of cancers, including head and neck cancers (HNC). In this study, we aimed to predict progression-free survival (PFS) in patients with HNC utilizing PET-, CT-, and multiple PET-CT fusion-based radiomic features. A total of 320 patients with HNC from 6 centers were included in this study, along with their tumor type and PET and CT images. We fused each patient’s PET and CT images using multiple multi-scale, saliency, subspace, and hybrid-based fusion methods. Therefore, we added 18 new image sets to the previous PET and CT sets. Using the SERA package (in agreement with IBSI guidelines), we extracted 104 textural and moment-invariant radiomic features, and the most relevant ones were selected via univariate C-index. Subsequently, a model was constructed using Cox proportional hazard (Coxph) machine learning algorithm. Our model was 6-fold cross-validated (leave-one-center-out cross-validation), and 100 bootstrapping was performed. Overall, the GTF fusion model had the best performance among all models (C-index ± SD = 0.66 ± 0.09). Following GTF, models of CBF, FPDE, HMSD, and TIF performed the best (C-indices ± SD of 0.65 ±0.08, 0.65 ±0.08, 0.65 ±0.09, and 0.65 ±0.08, respectively). In conclusion, our findings showed that the GTF fusion method is outstandingly better at creating PET-CT images that provide the best predictive performance for patients’ PFS.