A radio-pathological fusion model for predicting PD-L1 expression and immunotherapy response in non-small cell lung cancer.

OBJECTIVE: This study aims to construct a multimodal fusion model (FM) based on CT and hematoxylin and eosin (H&E) stained slices to predict the PD-L1 expression in non-small cell lung cancer (NSCLC) and to explore its additional value in predicting the prognosis of immunotherapy. MATERIALS AND METHODS: A retrospective analysis was conducted of 328 NSCLC patients with available PD-L1 immunohistochemical results. They were randomly divided into a training set, a validation set, and a test set in a 4:1:1 ratio. Radiomics and pathological models were constructed based on CT images and H&E slides, respectively, to predict PD-L1 expression, and then a radio-pathological FM was established. Then, the radio-pathological FM was used to generate predictive scores for an independent NSCLC immunotherapy survival validation cohort. RESULTS: A total of 55.5% (182/328) of patients were PD-L1 positive and included in the PD-L1 prediction cohort. Compared to the single-modality model, the radio-pathological FM achieved the highest predictive performance, with AUCs of 0.90, 0.80, and 0.73 across the three subsets, respectively. In the survival validation cohort, patients in the high-score group had significantly better progression-free survival (PFS) and overall survival than those in the low-score group. Furthermore, the FM score was an independent predictor of PFS. When combined with clinical factors, its C-index for predicting PFS was 0.74 (95% CI: 0.665-0.809). CONCLUSION: For the first time, a radio-pathological FM was constructed to predict PD-L1 expression in NSCLC. The study also demonstrated the model's potential for predicting patient prognosis under immunotherapy. CRITICAL RELEVANCE STATEMENT: This first fusion model combining CT radiomics and hematoxylin and eosin (H&E) deep learning non-invasively predicts programmed death-ligand 1 (PD-L1) and immunotherapy response in non-small cell lung cancer (NSCLC). KEY POINTS: The fusion model can accurately predict programmed death-ligand 1 (PD-L1) and immunotherapy outcomes in non-small cell lung cancer (NSCLC). The fusion model outperformed either single-modality model in distinguishing PD-L1-positive. Potential to reduce PD-L1 immunohistochemical testing and support precision immunotherapy decisions.