Construction and validation of a CT-based radiomics-deep learning signature for non-invasive prediction of PD-L1 expression and immunotherapy outcomes in non-small cell lung cancer.

Background: Non-small cell lung cancer (NSCLC) accounts for approximately 85% of all lung cancer cases. Although programmed death ligand-1 (PD-L1) immune checkpoint inhibitors (ICIs) have become a standard for advanced NSCLC; however, only 20-40% of patients achieve an objective response. For non-responders, ineffective treatment not only carries risks without benefits but also cause unnecessary consumption of medical resources. Currently, PD-L1 expression assays depend on invasive tissue biopsies, a method limited by sampling bias, etc. Therefore, there is an urgent need to develop non-invasive predictive tools. Alternative approaches such as liquid biopsies or positron emission tomography/computed tomography (PET/CT) models are being explored. However, CT, due to its low cost, high accessibility, and routine application in lung cancer diagnosis and management, serves as an ideal choice for developing non-invasive predictive tools. This study aims to derive a radiomics-deep learning signature (RADLsig) from CT images to predict PD-L1 expression in NSCLC and to further evaluate its utility in predicting clinical outcomes of immunotherapy. Methods: This study retrospectively included 804 patients with pathologically confirmed NSCLC who underwent baseline chest CT scans. After applying inclusion and exclusion criteria, 531 patients who underwent immunohistochemistry (IHC) for PD-L1 expression were randomly divided into a training set (n=424) and a validation set (n=107) in an 8:2 ratio to develop and validate the radiomics signature (RAsig), a deep learning signature (DLsig), and their fused signature (RADLsig) based on pre-treatment CT images. Radiomic features were extracted from manually delineated three-dimensional volumes of interest using the PyRadiomics platform. The response predictive performance of the RADLsig was validated in an independent immunotherapy cohort (n=145) consisting of patients who received PD-L1 checkpoint inhibitor immunotherapy. The primary efficacy endpoint was defined as objective response evaluated according to the immune-related Response Evaluation Criteria in Solid Tumors (iRECIST) criteria after three cycles of treatment. Additionally, in a The Cancer Imaging Archive (TCIA) cohort (n=128) containing matched CT and single-cell RNA sequencing data, we evaluated the correlation between RADLsig and standardized CD274 expression levels. Results: The study cohort comprised 804 NSCLC patients. The cohort used for model development (n=531) was predominantly adenocarcinomas (91.8%) with early-stage disease. The independent immunotherapy validation cohort [n=145, objective response rate (ORR) =51.7%], 49.0% were adenocarcinoma. For predicting PD-L1 expression, RADLsig achieved the highest area under the receiver operating characteristic (ROC) curve (AUC) [0.954; 95% confidence interval (CI): 0.901-0.986], significantly outperforming RAsig and DLsig (P<0.001 and P=0.03). Patients predicted as PD-L1 positive by RADLsig had a significantly higher response rate (64.4% Conclusions: RADLsig from CT images, preliminarily demonstrates the potential to predict PD-L1 expression status and response to immunotherapy in NSCLC patients Holding promise as a non-invasive auxiliary tool for patient selection in immunotherapy and may provide a reference for advancing individualized precision treatment in clinical practice.