In Academic radiology
RATIONALE AND OBJECTIVES : To develop a combined model incorporating the clinical and PET features for identifying patients with diffuse large B-cell lymphoma (DLBCL) at high risk of progression or relapse after first-line therapy, compared to International Prognostic Index (IPI) and Deauville score (DS) assessment.
MATERIALS AND METHODS : 271 18F-FDG PET images with DLBCL were retrospectively collected and randomly divided into the training (n=190) and test dataset (n=81). All visible lesions were annotated. Baseline, end-of-treatment (EoT), and delta PET radiomics features were extracted. IPI model, the baseline clinical model group (MG), DS model, the combined clinical MG, the PET-based radiomics MG, and the combined MG were constructed to predict 2-year time to progression (2Y-TTP). For each MG, the cross-combination method was performed to generate 1680 candidate models based on three normalization methods, 20 features, 4 feature-selection methods, and 7 classifiers. The model achieving the highest AUC was selected as the best for each MG. Cox regression analysis was further performed.
RESULTS : In the test set, the best combined model showed better discriminative power compared to IPI model, the best baseline clinical model, DS model, the best combined clinical model, and the best PET-based radiomics model (AUC 0.898 vs. 0.584, 0.695, 0.756, 0.824, 0.832; p < 0.001, 0.014, 0.018, 0.152, 0.042, respectively). The combined model was superior to other models for progression-free-survival prediction (C-index: 0.853 vs. 0.568, 0.666, 0.753, 0.808, 0.814, respectively).
CONCLUSION : A combined model for identifying patients at high risk of progression or relapse after first-line therapy was constructed, superior to IPI and DS assessment.
Cui Yingpu, Jiang Yongluo, Deng Xi, Long Wen, Liu Baocong, Fan Wei, Li Yinghe, Zhang Xu
2022-Nov-25
(18)F-FDG PET/CT, Diffuse large B cell lymphoma, Machine learning, Prognosis, Radiomics
