Deep learning-based PET/CT mixture-of-experts model for relapse risk stratification in relapsed/refractory classical hodgkin lymphoma: a multicenter study.

BACKGROUND: Relapsed/refractory classical Hodgkin lymphoma (R/R cHL) remains clinically challenging due to substantial heterogeneity in relapse risk. Reliable, non-invasive tools for improved relapse risk stratification are urgently needed to guide individualized therapeutic strategies. METHODS: In this multicenter retrospective study, 161 patients with R/R cHL from five institutions were included (training cohort: n = 102; validation cohort: n = 59). Clinical and metabolic covariates were assessed at the time of relapsed/refractory disease and baseline 18F-FDG PET/CT before salvage treatment. We developed a deep learning-based Mixture-of-Experts (MoE) framework that integrates four medical foundation models (PET-Diffusion, SAM-Med2D, MedCLIP, and RadFM) to derive a quantitative imaging biomarker (MoEScore) from baseline 18F-FDG PET/CT. A multiparametric model incorporating MoEScore and independent clinical/metabolic predictors was constructed and evaluated using discrimination, calibration, and clinical utility analyses. Model interpretability was assessed using attention visualization, ablation analysis, and pathological correlation. RESULTS: MoEScore demonstrated predictive performance (AUC: 0.861 in training; 0.783 in validation) and remained independently associated with relapse (HR = 11.18, 95% CI: 2.48-50.45; P = 0.002). The multiparametric model achieved a C-index of 0.785 in the training cohort and 0.717 in the validation cohort, compared with 0.598-0.752 for the clinical and clinical-metabolic models. MoEScore consistently stratified relapse risk across both relapsed and refractory subgroups. Exploratory interpretability analyses indicated that model saliency was predominantly localized to metabolically active lesion regions and suggested a greater relative contribution of PET than CT. MoEScore distributions were broadly consistent with known histopathological subtype patterns, supporting further biological evaluation. CONCLUSIONS: This study presents an exploratory, non-invasive deep learning framework for relapse risk stratification in R/R cHL. By integrating multimodal imaging and expert-level representations, the MoE model may capture tumor heterogeneity beyond conventional metrics and may help inform risk-adapted therapeutic strategies after further validation.