FDG-PET Radiomics Outperforms Conventional Biomarkers for Treatment Response Prediction In High-Risk Hodgkin Lymphoma: Evidence from the AHOD0831 Trial
Abstract
Purpose
Quantitative PET imaging biomarkers such as metabolic tumor volume (MTV) and maximum standardized uptake value (SUVmax) have demonstrated prognostic value in lymphoma; however, these metrics fail to characterize intratumor heterogeneity. This study aimed to assess whether PET radiomics can serve as a novel quantitative biomarker to predict treatment response in patients with high-risk Hodgkin lymphoma (HL).
Methods
A total of 115 PET/CT scans from patients enrolled in the AHOD08331 high-risk Hodgkin lymphoma trial were included in this analysis. From predefined metabolic tumor volumes on baseline PET scans, resampled to isotropic 4 mm voxels, 93 first- and second-order radiomic features were extracted using a fixed bin width between 0.1 and 0.3. All radiomic features were evaluated for association with treatment response using univariate statistical tests followed by LASSO-based feature selection. Logistic regression (LR) models were then constructed from the highest-ranked features and evaluated with receiver operating characteristic (ROC) curve analysis. The incremental benefit over baseline MTV- and SUVmax-based models was further quantified using decision curve analysis.
Results
LR models incorporating the top two selected RFs achieved a mean under the ROC curve (AUC) of 0.69, outperforming the baseline model using MTV and SUVmax (mean AUC=0.59, p<0.001). A hybrid model using the top RF feature and MTV also demonstrated improved performance (0.68, p<0.001). The top RF alone (AUC = 0.69) outperformed MTV alone (0.63) and SUVmax alone (AUC =0.50), both p<0.001. DCA revealed that the highest performing RF-based model provided a mean net benefit of 0.02 over the baseline model.
Conclusion
ROC analysis demonstrated that RF-based models consistently outperformed baseline models. DCA analysis confirmed a modest yet statistically significant net clinical benefit from integrating RFs into prediction models for treatment response in high-risk HL. These findings suggest that radiomics-augmented models could refine risk stratification and enable more personalized treatment decision.