Baseline PET Radiomics for Non-Invasive Risk Stratification In Hodgkin Lymphoma
Abstract
Purpose
Although most patients with Hodgkin lymphoma (HL) are cured with chemotherapy with or without immunotherapy, a subset fails to respond to first-line therapy. Baseline clinical factors alone do not reliably identify nonresponders. Baseline PET radiomic features may provide complementary prognostic information. We developed an automated radiomics pipeline to extract PET features and evaluate their association with clinical outcomes.
Methods
We identified 205 patients with HL with available tissue and baseline PET/CT imaging, comprising a testing cohort (n = 87) and an independent validation cohort (n = 118). A total of 1,316 radiomic features per patient were extracted using an automated PyRadiomics-based pipeline. In the testing cohort, feature selection was performed using a Lasso-Cox proportional hazards model to identify features with non-zero coefficients, followed by minimum redundancy maximum relevance ranking to prioritize features based on relevance and redundancy. Model performance was evaluated using progression-free survival analysis, time-dependent area under the curve (AUC), and concordance index (C-index). Feature-selection stability and generalizability were assessed using 5-fold cross-validation.
Results
Lasso-Cox selection identified 28 features (p < 0.0001), from which the top 9 features achieved strong prognostic performance (C-index = 0.77). Using a less stringent threshold (p < 0.001), 72 candidate features were identified, with the top 6 features achieving comparable discrimination (C-index = 0.77). Survival analyses demonstrated clear risk stratification, with low-risk patients exhibiting excellent long-term progression-free survival. Time-dependent AUC analysis showed improved performance for models using the p < 0.001 threshold. In 5-fold cross-validation, the 6-feature model demonstrated greater stability and superior performance (C-index = 0.78) compared with the 9-feature model (C-index = 0.71), with one feature consistently selected across both models.
Conclusion
Baseline PET/CT radiomic features show strong potential for non-invasive risk stratification in HL and may capture biologically relevant tumor microenvironment information. A six-feature model demonstrated improved stability and prognostic performance.