Background Immune checkpoint inhibitors (ICIs) have revolutionized the treatment of non-small cell lung cancer (NSCLC), benefiting 20–30% of patients. The current clinical standard for initiating ICI therapy is the assessment of Programmed Death-Ligand 1 (PD-L1) status via immunohistochemistry (IHC) on biopsy specimens. However, this invasive procedure presents risks and limitations, highlighting the need for a non-invasive alternative. This issue is critical as it affects patient outcomes and treatment accessibility. The challenge lies in accurately predicting PD-L1 expression using non-invasive methods, a task where naïve approaches often fail due to the complex nature of medical imaging data.

Methods This study retrospectively analyzed two cohorts of stage IV metastatic NSCLC patients undergoing immunotherapy, totaling 1080 individuals. Cohort 1 (n=746) was used for discovery, training (n=298), tuning, internal validation (n=75), and testing (n=373). Cohort 2 (n=334) served as an external validation set. The SCENT model was trained to predict PD-L1 status and stratify patients by progression-free survival (PFS) and overall survival (OS). Compared to traditional 2D, 2.5D, and 3D models, as well as radiomics and clinical predictors. We proposed a Scalable Ensemble Transformer (SCENT), a deep learning model, can predict PD-L1 expression from chest computed tomography (CT) scans in patients with metastatic NSCLC, thereby reducing the need for invasive biopsy procedures. Compared to prior studies, our approach differs by integrating multi-scale features from CT images, enhancing prediction accuracy and robustness, on large training and testing sets.

Results SCENT achieved superior performance in predicting PD-L1 status with specificity (81.59%), sensitivity (82.14%), and area under the curve (AUC; 80.50%). The model demonstrated robustness across varying training set sizes, achieving an AUC of 82.1%. In our evaluation, we compared PD-L1 status predictions (≥50% or <50%) using our deep learning model against other models, including radiomics, 2D, 2.5D and 3D models. SCENT’s predictions for OS and PFS were comparable to those derived from IHC-based PD-L1 status, validating its potential as a non-invasive diagnostic tool.

Conclusions The SCENT model represents a significant advancement in the non-invasive prediction of PD-L1 expression in NSCLC, offering a viable alternative to traditional biopsy methods. This innovation could streamline immunotherapy selection, making treatments more accessible and personalized. Future work will focus on refining the model and exploring its application in other cancer types.