18F FDG PET Radiomics features result in more accurate prediction of outcome for DLBCL patients than currently used IPI score

Introduction: Up to one-third of newly diagnosed diffuse large B-cell lymphoma (DLBCL) patients relapse or fail to achieve complete remission following first-line treatment. Adding PET radiomics features to currently used predictors may improve the identification of poor prognosis patients. The objective of this study was to externally validate the radiomics model developed in the HOVON-84 trial (Eertink et al, EJNMMI 2021) using datasets from other DLBCL studies within the PETRA database and to identify the optimal model to predict outcome in the PETRA database by combining radiomics features and clinical parameters.

Methods: 1092 newly diagnosed DLBCL patients with baseline ¹⁸F-FDG PET/CT scans with 2 year follow up in 7 studies in the PETRA database (https://petralymphoma.org) were included. Primary outcome was 2-year time to progression (TTP). 274 patients had progression within 2 years. Patients who died without progression within 2 years were not included in this analysis. Lesions were delineated using a fully automated preselection of ¹⁸F-FDG avid structures defined by a standardized Uptake Value (SUV) ≥ 4.0 and volume >3mL. Missed lesions were added and non-tumor regions were removed using the ACCURATE tool. 23 radiomics features were extracted: 5 conventional PET features (SUV_max, SUV_mean, SUV_peak, total lesion glycolysis and metabolic tumor volume (MTV)) and 18 dissemination features (number of lesions, 4 features quantifying distance (Cottereau et al, JNM 2020), 10 features quantifying the differences in intensity and 3 features quantifying the differences in volume between lesions) using RaCat software (Pfaehler et al, Plos One 2019). We tested the predictive value of 1) the currently used international prognostic index (IPI) score, 2) the HOVON-84 model (MTV, SUV_peak, the maximum distance between the largest lesion and any other lesion (Dmax_bulk), performance status and age) and 3) a model that combined radiomics features and individual components of the IPI score (PETRA model). For the PETRA model we applied logistic regression with backward feature selection. Survival curves were compared using the log rank test. Model performance was assessed using repeated cross-validation (5 folds, 2000 repeats) yielding the mean receiver-operator-characteristics curve integral (AUC). High- and low-risk groups were defined based on prevalence of events. Diagnostic performance was assessed using the positive predictive value.

Results: The categorical IPI score (model 1) yielded a cross-validated AUC (CV-AUC) of 0.64 ± 0.04 (AUC of individual datasets: 0.59-0.69; Figure 1A). The HOVON-84 trial (model 2) yielded a CV-AUC of 0.72 ± 0.04, this model was predictive of outcome in all 7 individual datasets within PETRA (AUC: 0.65-0.83). After backward feature selection, a combination of MTV, the maximum difference in intensity between the largest lesion and any other lesion (DSUVpeakbulk), DmaxBulk, lactate dehydrogenase levels and performance status resulted in the highest model performance for model 3 with a CV-AUC of 0.73 ± 0.04. This model was also predictive of outcome for all individual studies (AUC: 0.70-0.83; Figure 1B). High-risk patients according to model 3, had a 2-year TTP of 58.8% (95% CI: 53.2-64.9) which was significantly lower (p < 0.001) than the survival of low-risk patients (85.9% (95% CI: 83.6-88.3)). The positive predictive value increased from 28.1% (95% CI: 24.0-32.7) for IPI based high-risk patients (210 patients) to 46.0% (95% CI: 41.2-50.8) for PETRA based high-risk patients (274 patients).

Conclusions: In concordance with the HOVON-84 data, combining quantitative radiomics features extracted from baseline ¹⁸F-FDG PET/CT scans with components of the IPI score significantly improved identification of patients at risk of relapse when treated with standard first-line treatment regimens. Moreover, comparable radiomics features were selected in the HOVON-84 and PETRA prediction models.

• Download figure
• Open in new tab
• Download powerpoint