Deep Learning Analysis Based on Dual-energy CT-Derived Iodine Map for Predicting PD-L1 Expression in Gastric Cancer:...

PD-L1 expression is a critical biomarker in guiding immunotherapy for gastric cancer (GC). This study aims to investigate the value of deep learning analysis based on dual-energy CT-derived iodine map for predicting the level of PD-L1 expression in GC.

A total of 267 GC patients who underwent gastrectomy and preoperative dual-energy CT from multiple centers were prospectively enrolled and categorized into training (TC, n=143), internal validation (IVC, n=60), and external validation cohort (EVC, n=64). A 50-layer Residual Network was used to extract deep learning (DL) features from tumor volumes of interest on the iodine map. Machine learning was employed to develop the DL feature signature model (DFSigM). Multivariable logistic regression was used to screen PD-L1-related clinical characteristics, then a clinical model and a DL-clinical fusion model were also built. Model performance was evaluated based on discrimination, calibration, and clinical utility. Model interpretability was achieved through SHAP and Grad-CAM.

Following feature selection, 12 key DL features were identified and utilized to construct DFSigM. DFSigM achieved AUC values of 0.854 in TC, 0.836 in IVC, and 0.818 in EVC, outperforming the clinical model (AUCs of 0.785, 0.720, and 0.695), while comparable to the fusion model (AUCs of 0.858, 0.828, and 0.833). DFSigM provided a high net clinical benefit across a wide range of threshold probabilities, and also demonstrated good agreement between the predicted and actual probabilities. SHAP and Grad-CAM visualized the decision-making process of the model.

A deep learning model based on iodine map has been proven to be a valuable, reliable, and interpretable tool for non-invasive prediction of PD-L1 expression in GC.