Objective  To develop a ¹⁸F-fluorodeoxyglucose (FDG) PET/CT radiomics model based on stack learning for predicting the treatment response of patients with diffuse large B-cell lymphoma (DLBCL).

Methods  Retrospective analysis was conducted on 82 patients with DLBCL (44 males and 38 females, with a mean age of (57.9±12.8) years) who underwent ¹⁸F-FDG PET/CT at Changzhi People′s Hospital between March 2021 and June 2024. All patients were randomly divided into the training set (n=61) and the test set (n=21) in a 3∶1 ratio. Based on the Lugano 2014 classification system′s Deauville score, patients were categorized into the complete response group and non-complete response group. Regions of interest (ROI) on pre-treatment ¹⁸F-FDG PET images were pre-segmented using the thresholds standardized uptake value (SUV)≥4.0 and volume≥3 ml, followed by manual adjustments to obtain 297 radiomic features. Radiomic scores (RS) were calculated using the least absolute shrinkage and selection operator (LASSO). By integrating clinical characteristics and RS of DLBCL, a stacking learning model was developed using four base machine learning models (gradient boosting machines, random forest, naive Bayes, and neural network) to predict the treatment response of patients with DLBCL. Measurement data were compared between the two groups using either Student′s t-test or Mann-Whitney U test, whereas counting data were analyzed using the χ² test. The predictive performance of the models was evaluated by metrics, including the area under curve (AUC) of the receiver operating characteristic curve, accuracy, F1 score, precision, and recall. Differences in AUC among the models were compared using the Delong test.

Results  Of the 82 patients, 32 patients achieved complete response, whereas 50 patients were classified as non-complete response. In the training set, significant differences were observed in the presence of B symptoms, international prognostic index (IPI) scores, and bulky disease between the complete response and non-complete response groups (χ²=4.462–8.509, all P<0.05). LASSO analysis identified nine radiomic features related to treatment response of DLBCL training set patients for model construction. Among the five predictive models combining clinical features and RS, the stacking learning model demonstrated the best performance, achieving an AUC of 0.85 and an accuracy of 0.86 in the test set.

Conclusion  A stacking learning model integrating ¹⁸F-FDG PET/CT radiomics and clinical features improves risk stratification in patients with DLBCL.