2750 - Investigating PET/CT Radiomics for subtype Characterization and Treatment Response Prediction in Hodgkin Lymphoma

Purpose/Objective(s): Although most patients with Hodgkin lymphoma (HL) are effectively cured by chemotherapy with or without immunotherapy, many will not respond to primary treatment and baseline clinical features do not adequately predict nonresponders. We developed an automated pipeline to extract PET radiomic features, investigate associations with clinical outcomes, and identify unsupervised radiomic phenotypes reflecting underlying tumor biology.

Materials/Methods: We identified 204 HL patients with baseline PET/CT imaging. Radiomic features (1316 per patient) were extracted using an automated PyRadiomics-based pipeline. In a testing cohort (n=87), feature selection used Lasso-Cox modeling and mRMR ranking, with models evaluated by progression-free survival, time-dependent AUC, and C-index using 5-fold cross-validation. For unsupervised analysis in the independent validation cohort (n=117), 166 features were retained after variance filtering and correlation-based reduction. PCA followed by K-means clustering (K=3) identified distinct radiomic phenotypes visualized by UMAP.

Results: Lasso-Cox selection yielded 28 features (p<0.0001), with the top 9 achieving a C-index of 0.77. At p<0.001, the top 6 features achieved comparable discrimination (C-index=0.77) with greater cross-validation stability (C-index=0.78). Survival analyses showed clear risk stratification with excellent long-term progression-free survival in low-risk patients. In 5-fold cross-validation, the top 6-feature model showed greater stability and improved performance (C-index=0.78) compared with the top 9-feature model (C-index=0.71). Unsupervised analysis identified three stable radiomic phenotypes: Cluster 1 with irregular lobulated margins and asymmetric intensity profiles; Cluster 2 with compact, smooth, homogeneous tumors; and Cluster 3 with high intratumoral textural heterogeneity and complex gray-level distributions.

Conclusion: Baseline PET/CT radiomic features demonstrate strong potential for non-invasive risk stratification in HL and may capture biologically meaningful information related to the tumor microenvironment. Unsupervised phenotyping revealed three biologically distinct subtypes differing in textural heterogeneity, morphological complexity, and intensity distribution, potentially reflecting differences in tumor microenvironmental composition and growth patterns. A model using six radiomic features provided superior stability and prognostic performance, supporting its use in downstream validation and future integration with biological correlates. Additionally, we envision our model may inform patients who may derive benefit from combined systemic and radiation therapy based on baseline PET features.