[18F]FDG PET radiomics score generated by cross-combination approach for treatment response and prognosis prediction in primary gastrointestinal diffuse large B-cell lymphoma patients

Abstract

Objectives We investigated the value of using a machine learning cross-combination approach to construct a PET radiomics score (RadScore) for predicting the early treatment response and prognosis of patients with primary gastrointestinal diffuse large B-cell lymphoma (PGI-DLBCL) treated with the R-CHOP-like regimen. Methods We conducted a retrospective analysis on 108 PGI-DLBCL patients diagnosed between November 2016 and December 2021. Seven machine learning models were used to generate 49 feature selection-classification candidates, and the optimal candidate was selected to create RadScore. Logistic regression identified risk factors, and a radiomics nomogram combining RadScore with selected risk factors was constructed. The model was evaluated using calibration curves and decision curve analysis (DCA). Results A total of 111 radiomics features were extracted, and 19 features with strong predictive performance were used to generate RadScore. Logistic regression analysis in the training cohort identified elevated lactate dehydrogenase (LDH) level, intestinal involvement, and total lesion glycolysis (TLG) as independent risk factors for predicting early treatment response. The multi-parameter model incorporating RadScore, clinical risk factors, and metabolic factors showed good performance (training cohort AUC: 0.860; validation cohort AUC: 0.902). The RadScore is capable of effectively stratifying patients' progression-free survival (PFS) and overall survival (OS). Conclusions The machine learning-based RadScore can predict the survival of PGI-DLBCL patients. When combined with clinical risk factors and metabolic factors, it forms a combinatorial model suitable for predicting early treatment response to R-CHOP-like chemotherapy regimens.