Enhanced Lymphoma Subtype Classification and Prognosis Using Machine Learning with 18F-FDG PET/CT Radiomics: Beyond SUVmax

Abstract Background This study explored a machine learning approach using 18 F-FDG PET/CT as a non-invasive alternative to biopsy, incorporating tumor-to-liver ratio (TLR) PET radiomics, and performed survival analysis to improve lymphoma management. Methods In this cohort study, baseline 18 F-FDG PET/CT scans of newly diagnosed, histologically confirmed lymphoma patients were analyzed. Lesions were segmented using 3D Slicer, and radiomic features were extracted and normalized by tumor-to-liver ratios. Patient-level features were used to train three machine learning models (XGBoost, AdaBoost, Logistic Regression) using nested cross-validation with SMOTE for class balancing. A model based on SUV max metrics served as baseline. Radiomic features were also evaluated for correlation with 3- and 5-year survival using the Mann-Whitney U test. Results A total of 156 lymphoma patients were analyzed, with 2,076 lesions segmented and 200 radiomic features extracted. For subtype classification, AdaBoost achieved the highest AUC for Diffuse Large B-cell (DLBCL) (0.863, accuracy 0.742), while XGBoost performed best for High-Grade Non-Hodgkin lymphoma (NHL) (AUC 0.825, accuracy 0.735) and Nodular Sclerosis Hodgkin Lymphoma (NS-HL) (AUC 0.827, accuracy 0.832). Logistic regression showed the best results for Classical Hodgkin Lymphoma (C-HL) (AUC 0.849, accuracy 0.775). The SUV max -based model (LR-SUV_MAX) consistently underperformed (AUCs: C-HL 0.630, High-Grade NHL 0.700, NS-HL 0.638, DLBCL 0.664), with all differences being statistically significant ( p < 0.001). Radiomic and clinical features including SUV-GLSZM small area emphasis ( p = 0.0019), age ( p = 0.0002), and spleen involvement ( p = 0.0014) were significantly associated with 3- and 5-year overall survival in 110 and 74 patients, respectively. Conclusion Radiomic features combined with machine learning significantly improve lymphoma subtype classification over SUV max alone and show potential for predicting patient survival.