Radiomics reflecting both tumor and host features improves outcome prediction in follicular lymphoma
Résumé
Introduction: To date, several indices widely based on simple clinical or biologic parameters have been proposed to refine prognosis of follicular lymphoma (FL). The prognostic value of 18F-FDG PET/CT parameters such as Total Metabolic Tumor Volume (TMTV) remains controversial. Here, we explored the prognostic impact of additional features obtained from 18F-FDG PET/CT images in patients included in the phase III RELEVANCE trial (Morschhauser, NEJM 2018, JCO 2022), which compared rituximab-chemotherapy (R-chemo) with rituximab-lenalidomide (R2) in patients with previously untreated, high tumor burden FL.
Methods: Baseline 18F-FDG PET/CT scans and clinical information (ECOG, age, Ann Arbor stage, and FLIPI) were available for 351 follicular lymphoma patients. Lesions were segmented semi-automatically by expert physicians on the PET/CT scans. Deep learning tools (TotalSegmentator and MOOSE) were used to automatically segment organs from PET-registered CT scans. A total of 7437 PET and CT features were calculated, including tumor radiomics from segmented lesions and host radiomics from segmented organs and correlated to PFS and OS. To select predictive features, a permutation test was used to ensure that less than one false positive was selected. Highly correlated features were dropped to reduce feature redundancy and only features significantly predictive of both PFS and OS were selected. Finally, a Cox model was trained and evaluated in a 10x10 nested cross-validation with feature selection and hyperparameters tuning performed in the inner loop. Averaged time-dependent ROC-AUC (tAUC) was used to assess the prognostic value of the different features and models. Three models with different feature sets were built: basic (clinical features and TMTV), tumor (clinical features, TMTV, and tumor radiomics), and global (clinical features, TMTV, tumor radiomics, and host radiomics).
Results: Median number of selected tumor features was 5, reflecting tumor metabolic activity, and tissue densities measured on CT in lesion surroundings. They had an average univariate tAUC of 0.56 ± 0.03 for PFS and 0.59 ± 0.01 for OS. Median number of selected organ features was 2 with an averaged tAUC of 0.56 ± 0.04 for PFS and 0.60±0.06 for OS. Selected features reflected FDG uptake magnitude in liver, lung and kidney density. The basic model reached a tAUC of 0.58±0.04 for PFS and 0.65±0.05 for OS. The tumor model led to tAUC of 0.59 ± 0.04 for PFS and 0.67 ± 0.06 for OS. The global model yielded to a tAUC of 0.63 ± 0.04 for PFS and 0.72 ± 0.05 for OS. Global model was significantly better than clinical on both PFS and OS (p < 0.01) while tumor model was significantly better than basic model on PFS (p < 0.01) but not on OS (p < 0.27).
Conclusions: Our study suggests that radiomics features complementary to TMTV derived from baseline 18F-FDG PET/CT scans can improve outcome prediction for follicular lymphoma patients.