Cross-Regional Radiomic Consistency In Squamous Cell Carcinoma: Common PET/CT Features and Their Prognostic Value
Abstract
Purpose
Radiomics enables quantitative characterization of tumors from PET and CT imaging. When anatomic-region-specific cancer data are limited, transfer and semi-supervised learning can improve prediction by leveraging data from different regions; however, the basis of such transferability remains underexplored. Squamous cell carcinoma (SCC) occurs in both lung and head and neck (HN) cancers and shares a common histopathological origin, suggesting potential radiomic commonalities. This study identifies shared PET and CT-derived radiomic features (RFs) between lung and HN SCC, and evaluates their prognostic value for survival prediction.
Methods
PET and CT images from 42 lung-SCC patients and 42 HN-SCC patients were randomly selected from multicenter datasets, including TCIA Radiogenomics, TCGA, BC-Cancer, and HECKTOR challenge. PET images were SUV-normalized, and CT images were intensity-clipped before analysis. Tumor regions of interest were used to extract 107 RFs by PyRadiomics, including first-order, shape-based, and texture features. Feature distributions were compared between cancer types within each modality using the Mann-Whitney-U-test. Survival outcome prediction (good vs. poor) was performed using 11 classifiers for HN-CT, lung-CT, HN-PET, and lung-PET. Feature importance for each of the four prediction tasks was then assessed using SHapley Additive exPlanations analysis (SHAP).
Results
Thirteen PET- and nine CT-RFs showed no significant distributional differences between lung- and HN-SCC (p>0.05). In PET imaging, three shared first-order features (mean, 90th percentile, and total energy) ranked among the top predictors of outcome, whereas only one CT feature (90th percentile) demonstrated comparable importance. For lung-SCC, the best predictive performance was achieved with AUC=0.68±0.17 with Multilayer Perceptron (MLP) for PET and 0.68±0.10 with Gradient Boosting for CT. In HN-SCC, CT achieved AUC=0.87±0.10 with MLP, while PET reached AUC=0.80±16 using Random Forest.
Conclusion
Shared PET-RFs across lung- and HN-SCC may support semi-supervised and transfer learning by enabling transferable modeling with unlabeled data, improving cross-cancer prognostic prediction.