FDG-PET Intensity Normalization Improves Radiomics-Based Survival Prediction in Patients with Oropharyngeal Cancer: A Comparison of the Standardized Uptake Value with Alternative Normalization Techniques

BACKGROUND AND PURPOSE: Despite the widespread research application of radiomics, there is a knowledge gap regarding the optimal voxel intensity normalization strategy for FDG-PET radiomics. We investigated the impact of 3 normalization strategies on the prognostic utility of individual radiomic features and machine learning models in patients with oropharyngeal squamous cell carcinoma (OPSCC).

MATERIALS AND METHODS: We included n=330 (overall survival [OS] study group), n=335 (progression-free survival [PFS] study group), and n=309 (locoregional progression [LRP] study group) patients with OPSCC. Three FDG-PET intensity normalization strategies were applied: the conventional body weight–corrected standardized uptake value (SUV), and standardized uptake ratios to the lentiform nucleus and to the cerebellum. The raw PET voxel intensities were also analyzed. To quantify and compare the features’ association with oncologic outcome, we fitted univariate Cox regression models, calculated Harrell concordance index (C-index), and fitted random survival forest (RSF) machine learning algorithms.

RESULTS: All normalization strategies tended to improve the prognostic value of radiomic features. Features from lentiform nucleus–normalized PET demonstrated the highest prognostic improvement, with n=750/1037, n=809/1037, and n=652/1037 primary tumor features attaining a significant association with OS, PFS, and LRP, respectively, compared with n=0, n=211, and n=1 SUV-based PET features, respectively. The median C-index of lentiform nucleus–normalized PET features was 0.64, 0.61, and 0.62 for OS, PFS, and LRP, respectively, while SUV-based PET features reached 0.59, 0.58, and 0.60, respectively. The best performing lentiform nucleus-normalization RSF model significantly outperformed the raw PET RSF model in predicting OS (C-index = 0.66 versus C-index = 0.57; P = .019), with model comparisons for PFS and LRP approaching statistical significance (P = .053 and P = .084, respectively). In contrast, the best performing SUV-based RSF models were not significantly different from raw PET models.

CONCLUSIONS: Normalizing PET intensities, especially to the lentiform nucleus, improves the prognostic performance of individual radiomic features and machine learning models in predicting oncologic outcome.