Deep learning–based PET/CT radiomics-clinical model for prognostic stratification and immunotherapy exploration in natural killer/T-cell lymphoma.

Abstract

7072 Background: Natural killer/T-cell lymphoma (NKTCL) is a highly aggressive malignancy with substantial heterogeneity. The value of deep learning on PET/CT-based radiomics for risk stratification in NKTCL remains to be defined. Methods: In this retrospective multicenter study, 425 consecutive NKTCL patients from four hospitals were randomly assigned to training or validation cohorts. Radiomics features extracted from pretreatment PET/CT images were used to construct prognostic models with multiple deep learning algorithms. The best model was integrated with established clinical prognostic factors, and its performance in informing treatment decisions was tested in another independent immunochemotherapy cohort. Results: The XGBoost-based radiomics model achieved the highest prognostic accuracy for progression-free survival (PFS) and overall survival (OS) across cohorts. Integration with clinical factors yielded the Radiomic-Clinical Prognostic Model (RCPM), which further improved prediction (training cohort AUC: PFS 0.977, OS 0.946; validation cohort AUC: PFS 0.834, OS 0.814) and effectively stratified patients into high- and low-risk groups with distinct PFS and OS (all P < 0.001). In another exploratory treatment cohort, high-risk patients derived significant benefit from PD-1 inhibitor therapy (3-year PFS: 67.7% vs. 21.7%, 3-year OS: 81.7% vs. 39.6%; both P < 0.001), whereas low-risk patients did not. Conclusions: In conclusion, the RCPM demonstrated strong prognostic value and may help identify patients who could potentially benefit from immunotherapy in NKTCL.