Metabolic heterogeneity‐based radiomic model predicts treatment outcomes of primary mediastinal B‐cell lymphoma patients in the <scp>IELSG37</scp> study

Early identification of patients at risk of failing first-line therapy remains a key challenge in primary mediastinal B-cell lymphoma (PMBCL). This study assessed whether positron emission tomography (PET) radiomics could predict treatment outcomes in PMBCL. Baseline PET/computed tomography (CT) scans from 501 patients enrolled in the International Extranodal Lymphoma Study Group 37 (IELSG37) trial and treated with rituximab- and doxorubicin-based immunochemotherapy were analysed. Functional PET parameters reflecting tumour burden (metabolic tumour volume, MTV), glucose consumption (total lesion glycolysis, TLG) and metabolic heterogeneity (MH) were calculated using a 25% maximum standard uptake value (SUVmax) segmentation threshold. Radiomic analysis was performed in 495 patients, extracting 107 features; 74 features uncorrelated with SUVmax and MTV were selected and analysed using a recursive-partitioning classification tree (Ctree) approach. Ctree identified grey-level run-length matrix run variance (GLRLM-RV), a marker of MH, as the most significant prognostic feature. Patients with low MH (GLRLM-RV <0.136) had a 5-year progression-free survival (PFS) of 93%, compared with 65% in those with high MH (p < 0.0001). Five-year overall survival (OS) was 96% vs. 80% respectively (p < 0.0001). A subsequent Ctree model integrating GLRLM-RV and TLG further improved prognostic stratification for PFS and OS (p < 0.0001), outperforming established prognostic indices (International Prognostic Index [IPI], revised-IPI and age-adjusted IPI).