CT Image-Based Radiomic Analysis for Detecting PD-L1 Expression Status in Bladder Cancer Patients

Rationale and Objectives: The role of Programmed death-ligand 1 (PD-L1) expression is crucial in guiding immunotherapy selection. This study aims to develop and evaluate a radiomic model, leveraging Computed Tomography (CT) imaging, with the objective of predicting PD-L1 expression status in patients afflicted with bladder cancer. Materials and Methods: The study encompassed 183 subjects diagnosed with histologically confirmed bladder cancer, among which the PD-L1(+) cohort constituted 60.1% of the total population. Stratified random sampling was utilized at a 7:3 ratio. We employed five diverse machine learning algorithms-Decision Tree, Random Forest, Linear Support Vector Classification, Support Vector Machine, and Logistic Regression-to establish radiomic models on the training dataset. These models endeavored to predict PD-L1 expression status premised on radiomic features derived from region-of-interest segmentation. Subsequent to this, the predictive performance of these models was examined on a validation set employing the receiver operating characteristic (ROC) curve. The DeLong test was utilized to contrast ROC curves, thereby pinpointing the model with superior predictive accuracy. Results: 16 features were chosen for the model construction. All five models revealed strong performance in the training set (AUC, 0.920-1) and commendable predictive ability in the validation set (AUC, 0.753-0.766). As per the DeLong test, no statistically significant disparities were observed among any of the models (P > 0.05) in the validation set. Additional verification through the calibration curve and decision curve analysis indicated that the Logistic Regression model exhibited extraordinary precision and practicality. Conclusion: Our machine learning model, grounded on radiomic features, demonstrated its proficiency in accurately distinguishing bladder cancer patients with high PD-L1 expression. Future research, incorporating more exhaustive datasets, could potentially augment the predictive efficiency of radiomic algorithms, thereby advancing their clinical utility.