整合生境放射组学与2.5D深度特征的非小细胞肺癌PD-L1 TPS≥50%无创预测

OBJECTIVE: To develop and evaluate a noninvasive imaging framework based on 18F-FDG PET/CT for estimating programmed death-ligand 1 (PD-L1) expression status in non-small cell lung cancer (NSCLC) by integrating habitat radiomics and 2.5D deep features.

MATERIALS AND METHODS: This retrospective two-center study included 224 patients with pathologically confirmed NSCLC who underwent pretreatment 18F-FDG PET/CT and PD-L1 immunohistochemistry. Patients from Hospital 1 were randomly assigned to a training cohort (n = 138) and a test cohort (n = 42), while patients from Hospital 2 served as an independent external validation cohort (n = 44). PD-L1 expression was dichotomized as low (TPS < 50%) or high (TPS ≥ 50%). Radiomics features were extracted from PET and CT images, and habitat subregions were generated using K-means clustering. 2.5D deep features were extracted using a ResNet-50-based feature-extraction strategy with maximum-intensity projections. Prediction models were constructed using support vector machine, XGBoost, logistic regression, random forest, and artificial neural network classifiers, and were evaluated using the area under the receiver operating characteristic curve with 95% confidence intervals, accuracy, sensitivity, specificity, precision, F1 score, and balanced accuracy. SHAP was applied to enhance the visualization and interpretability of the models.

RESULTS: Habitat radiomics demonstrated superior discriminative performance compared with conventional radiomics across cohorts, particularly in the external validation cohort (AUC = 0.840 vs 0.794). The 2.5D deep-feature model achieved the highest discriminative ability in external validation (AUC = 0.918), indicating favorable generalizability. The 2.5D deep-feature model achieved the highest external-validation AUC and F1 score among the representative models (AUC = 0.918; F1 score = 0.783). The combined model achieved the highest AUC in the test cohort (AUC = 0.894) and showed favorable sensitivity in external validation (0.769), although its external-validation AUC and F1 score were slightly lower than those of the 2.5D deep-feature model. SHAP analysis revealed that deep features and habitat radiomics features reflecting intratumoral heterogeneity were the primary contributors to model predictions, whereas conventional metabolic parameters showed limited impact.

CONCLUSION: In conclusion, the proposed 18F-FDG PET/CT-based framework integrating habitat radiomics and 2.5D deep features showed promising performance for noninvasive estimation of PD-L1 expression status in NSCLC. The combined model demonstrated balanced predictive performance across the test and external validation cohorts, suggesting its potential as a candidate imaging biomarker. Further validation in larger multicenter studies is warranted. However, further prospective validation incorporating immunotherapy-response and survival outcomes is required before clinical application.