Abstract P057

Baseline PET Radiomics for Risk Stratification in Classic Hodgkin Lymphoma

Purpose: To identify distinct subgroups of patients with classic Hodgkin lymphoma (cHL) using unsupervised clustering of PET/CT-derived radiomic features, and to evaluate their prognostic associations with freedom from progression (FFP).

Methods: Baseline 18FDG PET/CT images were processed using an automated segmentation and PyRadiomics-based pipeline, extracting 1,316 features. Each feature was log transformed, z-score standardized, with a redundancy reduction correlation coefficient filter applied. Principal component analysis and a Gaussian mixture model (GMM) were used to assign radiomic clusters. Freedom from progression (FFP) was measured using Kaplan–Meier. A multivariable Firth-penalized Cox regression model was constructed using cluster membership after adjustments for age, stage, and sex.

Results: We identified 197 patients with cHL diagnosed at Stanford University between 2015-2024, with median age of 32 (range=6-79), 64.5% with early vs 35.5% with advanced stage, median IPS of 2, and median follow-up of 3.98 years. The GMM model partitioned into three radiomic clusters: RC1 (n=104; 53%), RC2 (n=53; 27%), and RC3 (n=40; 20%). RC1 was defined by heterogeneous texture and low tumor volumes, RC2 by homogenous texture and low tumor volumes, and RC3 by heterogeneous texture and high tumor volumes. The 4-year FFP was 68.2%, 82.2%, and 48.4% for RC1, RC2, and RC3, respectively (P=0.015). In a multivariable Cox model adjusted for age, stage, and sex, both RC1 (HR 1.96, 95% CI 1.01–4.13, p = 0.046) and RC3 (HR 2.82, 95% CI 1.32–6.37, p = 0.007) had higher risk of a PFS event relative to RC2. Posterior probabilities from the GMM clustering provided each patient with a continuous probability of belonging to each radiomic cluster. In time-dependent survival analyses and receiver-operator curves, a clinical-radiomic risk model incorporating these GMM posterior probabilities and 5-fold cross-validation trials (500 iterations) yielded high area under curve performance (AUC=0.68 at 4 years), which outperformed total metabolic tumor volume alone (AUC=0.61).

Conclusion: Unsupervised clustering of PET/CT-derived radiomic features can reproducibly identify three prognostically distinct cHL subgroups. These clusters capture significant FFP differences of these subgroups that remain independent of established clinical covariates including age, stage, and sex, suggesting that radiomic phenotyping may provide independent and complementary prognostic value.