Assessing PD-L1 expression in non-small cell lung cancer and predicting responses to immune checkpoint inhibitors using deep learning on computed tomography images

Rationale: This study aimed to use computed tomography (CT) images to assess PD-L1 expression in non-small cell lung cancer (NSCLC) and predict response to immunotherapy. Methods: We retrospectively analyzed a PD-L1 expression dataset that consisted of 939 consecutive stage IIIB-IV NSCLC patients with pretreatment CT images. A deep convolutional neural network was trained and optimized with CT images from the training cohort (n = 750) and validation cohort (n = 93) to obtain a PD-L1 expression signature (PD-L1ES), which was evaluated using the test cohort (n = 96). Finally, a separate immunotherapy cohort (n = 94) was used to assess the prognostic value of PD-L1ES with respect to clinical outcome. Results: PD-L1ES was able to predict high PD-L1 expression (PD-L1 >= 50%) with areas under the receiver operating characteristic curve (AUC) of 0.78 (95% confidence interval (CI): 0.75 similar to 0.80), 0.71 (95% CI: 0.59 similar to 0.81), and 0.76 (95% CI: 0.66 similar to 0.85) in the training, validation, and test cohorts, respectively. In patients treated with anti-PD-1 antibody, low PD-L1ES was associated with improved progression-free survival (PFS) (median PFS 363 days in low score group vs 183 days in high score group; hazard ratio [HR]: 2.57, 95% CI: 1.22 similar to 5.44; P = 0.010). Additionally, when PD-L1ES was combined with a clinical model that was trained using age, sex, smoking history and family history of malignancy, the response to immunotherapy could be better predicted compared to either PD-L1ES or the clinical model alone. Conclusions: The deep learning model provides a noninvasive method to predict high PD-L1 expression of NSCLC and to infer clinical outcomes in response to immunotherapy. Additionally, this deep learning model combined with clinical models demonstrated improved stratification capabilities.