Computed tomography texture analysis for prediction of PD-L1 expression in non small cell lung cancers

Objective: This study aims to evaluate the predictability of PD-L1 expression status as positive or negative and strong or weak positive based on radiomic features obtained from arterial phase contrast-enhanced computed tomography images in non-small cell lung cancers. Materials and Method: This retrospective and single-center study included 119 patients who were diagnosed histopathologically with non-small cell lung cancer (NSCLC), had arterial phase contrast-enhanced computed tomography (CT) images prior to treatment between January 2020 and February 2025, and had PD-L1 expression levels available in the database. For radiomics analysis, the largest diameter tumor was selected on CT images using the LIFEx software, and masses were manually marked in each sequence with semi-automatic segmentation and three-dimensional region of interest (3D ROI). A total of 127 radiomic features were extracted for each patient, including 28 shape-based, 41 first-order, and 58 second-order features. Radiomics segmentation consistency was evaluated using intraclass correlation analysis. For statistical analysis of radiomic data, independent sample t-test or Mann–Whitney U test was used for comparison of two independent groups according to distribution characteristics, and chi-square test was used for comparison of categorical variables. For diagnostic performance evaluation, ROC analysis was performed, and the optimal cut off point was determined using Youden's J index. Differences in ROC analysis performance were compared using the DeLong test. Logistic regression analysis was applied to develop a model for predicting PD-L1 groups, and classification accuracy was determined. The overall validity of the model was evaluated with the Omnibus Test, and model fit was reported with the Nagelkerke R Square value. The performance of the logistic regression model was also assessed by ROC analysis. For ease of clinical use, decision tree analysis was also performed, and classification accuracy was evaluated. In all statistical analyses, a p-value of <0.05 was considered statistically significant. Results: Our study demonstrated that radiomics models can achieve accuracy rates up to 74% (AUC: 0.78) in predicting positive or negative PD-L1 expression, and up to 68% (AUC: 0.744) in predicting strong positive or weak positive expression. Conclusion: Our study demonstrated that radiomic features obtained from arterial phase contrast-enhanced thoracic CT images can non-invasively predict PD-L1 expression in NSCLC patients with moderate accuracy.