Abstract
Background
This study aims to build a clinical factor model by incorporating clinical factors and metabolic parameters, as well as lesion imaging features from PET/CT images. Additionally, radiomics models are established based on PET-CT images to assess its capability in predicting PD-L1 expression in patients with non-small cell lung cancer.
Methods
After retrospective data collection, based on the clinical factor logistic regression results, a clinical factor model was constructed. The regions of interest (ROIs) for PET in radiomics were delineated using a semi-automatic method, while those for diagnosis CT were manually delineated. After extracting radiomic features, feature selection was performed using variance analysis, correlation analysis, and Gradient Boosting Decision Tree (GBDT). PET, diagnosis CT, and combined models were constructed. Predictive power was evaluated through ROC analysis comparing different models.
Result
In all 104 cases(mean age, 63.90years+/-8.99, 62males) were evaluated. The SUVmax in the PD-L1 positive group was higher than that in the negative group (P = 0.04), but both metabolic parameters and imaging features showed no correlation with PD-L1 expression. The radiomics models outperformed the clinical factor model (AUC = 0.712), yet the clinical factor model exhibited higher specificity than all radiomics models (Specificity = 0.765). The predictive performance of the PET model surpassed that of the diagnosis CT model (AUC: 0.838 vs 0.723). The combined model demonstrated enhanced predictive performance (AUC = 0.874).
Conclusion
The radiomics models perform better in predicting PD-L1 expression than the clinical factor model. The radiomics model combining PET and diagnosis CT exhibits the best predictive performance.