Baseline pet radiomics outperforms clinical risk scores in predicting primary mediastinal b‐cell lymphoma outcome: insights from the ielsg37 study

Introduction: A major challenge in treating primary mediastinal B-cell lymphoma (PMBCL) is the early identification of patients who are likely to fail first-line therapy. To address this issue, the present analysis explored the potential of using PET radiomics, a new technique for extracting quantitative parameters from PET images, to predict treatment outcomes in PMBCL patients. Methods: The study examined baseline PET/CT scans from 501 PMBCL patients treated with rituximab and doxorubicin-based immunochemotherapy regimens, with 342 of them receiving consolidation radiotherapy. Functional PET parameters indicating tumor burden (metabolic tumor volume, MTV), glucose consumption (total lesion glycolysis, TLG), and heterogeneity (area under curve of cumulative SUV-volume histogram, AUC-CSH) were calculated for each patient using a segmentation algorithm with a 25%-SUVmax threshold. In addition, 107 radiomics features (RF) were extracted from the baseline scans of 495 patients using the PyRadiomics package. 74 RFs uncorrelated with SUVmax and MTV were selected and then explored with a machine-learning approach based on the recursive-partitioning classification tree (CTree) method. Statistical analysis was performed using either Stata-17 or R software packages, as appropriate. Results: The CTree analysis identified a single RF with significant prognostic impact: the Grey-Level-Run- Length-Matrix run variance (GLRLMrv), which is a marker of metabolic heterogeneity (MH). Patients with low MH (GLRLMrv <0.137) had a 93% progression-free survival (PFS) rate at 5 years, while those with high MH had a 65% PFS rate (log-rank test, p < 0.0001). Overall survival (OS) rates were 96% and 80%, respectively (log-rank test, p < 0.0001). A subsequent CTree analysis was conducted, including GLRLMrv, as well as other dichotomized clinical variables and PET metrics that had a significant association with PFS (p < 0.05) at univariable analysis (TLG, MTV, SUVmax, SUVpeak, maximal lesion diameter stage, sex, LDH, and number of extra-nodal localizations). The CTree selection generated a new prognostic model based on the combination of GLRLMrv and TLG, which discriminated patients with different PFS and OS (p < 0.0001 for both). Concordance probability estimate showed a higher predictive accuracy (highest Harrell’s C value) of this radiomic model in comparison with the main international prognostic indices, namely IPI, revised-IPI and age-adjusted IPI. The better discriminatory power was also confirmed by its lowest Akaike’s information criterion. The research was funded in part by grants from the Swiss National Science Foundation (SNSF) – Project 32003B_146931, Krebsforschung Schweiz – Project KFS-2852-08-2011 and Cancer Research UK (C30423/A16247) The enrolment of Swiss patients was supported by the Swiss Group for Clinical Cancer Research (SAKK). This analysis was funded by a grant from the Swiss Cancer League KLS-5406-08-2021. Keywords: diagnostic and prognostic biomarkers, PET-CT No conflicts of interests pertinent to the abstract.