A proof of concept study of 18F-FDG PET/CT patient-level radiomics identify refractory/relapsed diffuse large B-cell lymphoma

Abstract

Abstract This study aimed to evaluate diffuse large B-cell lymphoma (DLBCL) patients who have refractory/relapsed disease and characterize the heterogeneity of DLBCL using patient-level radiomics analysis based on 18F-FDG PET/CT. A total of 132 patients diagnosed with DLBCL who underwent 18F-FDG PET/CT before receiving treatment were selected for the final study. Patient-level volumes of interests (VOI) were extracted from PET/CT images, and 328 radiomics features were extracted subsequently. 8 radiomics features were selected using the Least Absolute Shrinkage and Selection Operator (LASSO) algorithm to calculate the radiomics score (rad-score). Additionally, a total of 64 potential ML classifiers were generated based on 8 distinct supervised learning algorithms. The combined model that integrates rad-scores, clinical features and standard PET parameters demonstrates excellent performance; Specifically, ML models based on Naive Bayes have the greatest predicted values (AUC = 0.73). The patient-level radiomics features were subjected to unsupervised non-negative matrix factorization (NMF) clustering analysis to identify 3 radiomics subtypes. Cluster 1 exhibited a substantially higher prevalence of refractory/relapsed DLBCL compared to Clusters 2 and 3 (P < 0.05). Moreover, Cluster 1 showed a significantly higher frequency of advanced Ann Arbor stage, high international prognostic index, and bulk disease (all P < 0.05). In conclusion, Radiomics scores and radiomics subtypes derived from patient-level data offer significant predictive value and phenotypic information for patients with refractory/relapsed DLBCL.