Prediction of PD-L1 expression in gastric cancer using 18F-FDG PET/CT radiomics.

Abstract

Objective: To evaluate the predictive performance of radiomics models based on ¹⁸F-FDG PET/CT for determining programmed cell death ligand-1 status in treatment-naïve gastric cancer.

Methods: This retrospective study included 170 gastric cancer patients who underwent preoperative ¹⁸F-FDG PET/CT. Patients were randomly assigned to training (n = 119) and testing (n = 51) cohorts. PD-L1 positivity was defined as a combined positive score ≥5. Radiomic features were extracted from PET, CT, and fused PET/CT images and processed through a comprehensive pipeline with multiple methods employed for normalization, preprocessing, and feature selection. Each pipeline retained 20 radiomic features, which were subsequently evaluated with 10 different classifiers using feature subsets ranging from 1 to 20 to identify the optimal radiomics model. Model performance was assessed by AUC, accuracy, and related metrics, with 5-fold cross-validation on the training set.

Results: The integrated PET/CT model achieved the best performance on the test set (AUC 0.750; accuracy 0.745). The CT-based model yielded a test AUC of 0.678 and accuracy of 0.762, while the PET-based model showed a test AUC of 0.653 and accuracy of 0.706. All three radiomics models outperformed the conventional metabolic-parameter model (AUC 0.578; accuracy 0.544). The optimal predictive pipeline varied across the different imaging modalities.

Conclusion: Integrated ¹⁸F-FDG PET/CT radiomics demonstrated superior potential for noninvasive prediction of PD-L1 expression in gastric adenocarcinoma compared with single-modality imaging or metabolic parameters, supporting its potential utility for guiding patient selection for immunotherapy, with further validation warranted.