Non-invasive PD-L1 prediction in NSCLC patients using 3D self-supervised deep learning and radiomics.

3067 Background: Non-small cell lung cancer (NSCLC) is the most common subtype of lung cancer for ~85% of all cases. Despite therapeutic advances, prognosis remains poor, especially in advanced stages. Immunotherapy has revolutionized NSCLC treatment, with immune checkpoint inhibitors (ICIs) targeting the programmed death-ligand 1 (PD-L1) pathway. While PD-L1 expression is typically measured via immunohistochemistry (IHC), predictive modeling using CT images could offer a non-invasive alternative to enhance patient stratification and treatment planning in hard-to-biopsy cases and tumor follow-up of clonal resistance. We propose a solution for non-invasive prediction of PD-L1 expression leveraging radiomics and AI. Methods: This multicentric retrospective study included NSCLC patients from five real-world data sources who underwent CT and biopsy. CT scans were quality-checked and annotated by imaging experts supervised by radiologists (> 5 years' experience) to delineate primary tumors. PD-L1 levels were obtained via IHC. The cohort was randomly split into training and test sets (80/20%), with 5-fold cross validation for model building and fine-tuning. Three methods—radiomics, deep learning, and deep radiomics—were proposed for binary PD-L1 prediction (cut-off > 1%) based on 3D lesion-centered patches. The radiomics pipeline included 3D feature extraction, standardization, dimensionality reduction and classifier selection. The deep learning approach used a self-supervised 3D network, pretrained on 2420 lung lesion patches from 751 CTs by minimizing dissimilarity between augmented pairs and then fine-tuned on the training set to predict PD-L1 expression. The deep radiomics method fused both methods via weighted averaging of predicted probabilities. Results: A total of 324 patients (41% women, 63 ± 10 years) with varying PD-L1 expression (63% with levels > 1%), were included. The deep radiomics approach achieved the highest performance, with AUCs of 75.9% ± 5.3% and 70.1%, and F1-scores of 78.1% ± 1.2% and 76.7% on the validation and test sets, respectively, with a per-instance processing time of 1.2 ± 1.3 seconds. Compared to the radiomics method, it improved AUC by 2.6% and 1.2%, and F1-score by 2.7% and 3% on the validation and test sets, respectively. Compared to the deep learning approach, it showed AUC gains of 0.3% and 2%, and F1-score gains of 11.3% and 18.4% on the validation and test sets, respectively. Conclusions: This study demonstrates the effectiveness of a novel 3D image-based approach combining radiomics and 3D self-supervised learning to predict PD-L1 expression in a heterogeneous NSCLC cohort using real-world data. The model executes in seconds and could be regulatory cleared and deployed in clinical practice as a medical device performing non-invasive PD-L1 expression identification from CT scans.