Abstract Background Mantle cell lymphoma (MCL) is a rare, biologically heterogeneous B-cell malignancy with highly variable outcomes. Existing prognostic tools are suboptimal. We developed an interpretable deep learning framework integrating baseline 18 FFDG PET/CT and electronic health record (EHR) data for individualized risk stratification. Methods In this multicenter study, 187 treatment-naïve MCL patients were analyzed. A mixture-of-experts (MoE) fusion network integrated multimodal representations from PET/CT and EHR data. Expert modules comprising vision encoders, radiomics extractors, and a medical language model were integrated through an attention-based gating mechanism to construct multimodal radiomic signatures (R-signatures) predictive of progression-free survival (PFS) and overall survival (OS). R-signatures were validated and incorporated with clinical and metabolic factors into multiparametric models. Deep learning model interpretability was evaluated using attention visualization, expert-level contributions and pathologic correlation. Results R-signatures robustly discriminated relapse (AUC = 0.893 training, 0.755 validation) and death (AUC = 0.804 and 0.844), and independently predicted adverse outcomes (PFS: HR = 27.70, P < 0.001; OS: HR = 6.86, P = 0.001). Multiparametric models integrating R-signatures with total lesion glycolysis, β2-microglobulin, WBC, and Ki-67 outperformed conventional indices (C-indices: PFS 0.892 training, 0.781 validation; OS 0.877 training, 0.862 validation). Time-dependent ROC analyses consistently showed AUCs approaching or exceeding 0.800. Calibration and decision curve analyses confirmed excellent agreement and superior clinical net benefit. Attention maps localized high-weighted regions to hypermetabolic tumor areas, with higher R-signature values in blastoid and pleomorphic variants versus classical histology ( P = 0.028 and P = 0.010). Conclusions This interpretable PET/CT-EHR fusion framework substantially improves prognostic precision in MCL, providing a noninvasive, clinically translatable tool for risk-adapted management.
Jiang et al. (2026) studied this question.