Radiomics with super-voxel segmentation improves HPV prediction accuracy significantly in oropharyngeal cancer

Oya Altinok,Matthew B. Schabath,Albert Guvenis

DOI: https://doi.org/10.1101/2024.12.23.24319468

2024-12-26

Abstract:Human papillomavirus (HPV) status has been shown to be prognostic among patients with oropharyngeal cancer (OPC); Specifically, patients with HPV-positive tumors often have a superior prognosis and response to treatment compared to HPV-negative tumors which may be attributed greater tumor heterogeneity. This study assessed if analyzing tumor subregions (i.e., habitats) using super-voxel segmentation can effectively capture intratumor heterogeneity and improve predicting HPV positivity compared to a conventional whole-tumor approach. Using publicly available data from The Cancer Imaging Archive (TCIA) of 192 patients (85% HPV positive) with OPC, we utilized radiomics to predict HPV status comparing a super-voxel segmentation approach and a whole tumor approach. For the subregion approach, the number of supervoxels (subregions) generated per patient varied based on tumor size (mean = 30 supervoxels/patient [SD = 10]). 18 radiomic features were extracted from each supervoxel based on gray-level frequency distribution and aggregated using variance to summarize heterogeneity. For the whole tumor approach, the same radiomic features were generated across the entire tumor without sub-segmentation. As such, 18 radiomic features were utilized to predict HPV status in both models. The dataset was divided into a training set (70%) and an independent test set (30%). An optimizable ensemble model based on a decision tree with GentleBoost was applied to both the subregion and the whole tumor models to predict HPV status from radiomics features. The proposed super-voxel-based approach yielded an AUC of 0.94 in the training set and 0.91 in the test set which outperformed whole tumor analysis (AUC of 0.77 and 0.75, respectively). These findings demonstrate the value of incorporating heterogeneity measures and super-voxel segmentation in oropharyngeal cancer radiomics, which enable a significantly more accurate prediction of the HPV Status.