Deep-learning analysis of CT imaging biomarker for PD-L1 expression to predict heterogeneous response to immune checkpoint inhibitors in non-small cell lung carcinoma.

Abstract

e21529 Background: Inter-tumoral genomic heterogeneity cause various immune status in tumor microenvironment, which may lead to indiscriminate response to immune checkpoint inhibitor (ICI) in patients with multiple lesions. Therefore, PD-L1 status from practically approachable, single lesion would not be always representative for immune status of whole metastatic lesions of a patient. To solve inter-tumoral heterogeneity issue before biopsy in clinic, we developed a deep-learning based CT biomarker for predicting PD-L1 status, then explored if the algorithm would also predict ICI response of multiple lesions in non-small cell lung carcinoma (NSCLC). Methods: Deep learning-based image analyzer was trained with CT images of NSCLC in Samsung Medical Center (SMC) (N = 104). Taking 3D patch of a lesion located by radiologists, 3D convolutional neural network was trained to predict PD-L1 (22C3 immunohistochemistry) tumor proportion score of each lung or lymph node lesion. The prediction model was validated using publicly available dataset (NSCLC Radiogenomics, N = 115). Finally, we applied the model to baseline CT who had multiple lesions (≥ 2) and also received ICI (SMC validation, N = 170). Tumor response was assessed based on RECIST 1.1, and discordant response was defined by best response of each lesion outside -10% ~ +10%. Results: Predicted PD-L1 score was positively correlated with real PD-L1 expression in NSCLC Radiogenomics (Pearson = 0.198, P= 0.0339). In SMC validation cohort, predicted PD-L1 score of each lesion was negatively correlated with ICI response of corresponding lesion (Pearson = -0.0941, P= 0.0325). Interestingly, 35 out of 170 (20.6%) patients showing discordant ICI response among lesions had worse progression free survival (PFS) (median PFS: 2.1m, 1.6m, 2.5m, and 18.7m in discordant response, concordant progress, stable, and regress, respectively, P< 0.001). Patients with discordant response had significantly wide-ranged predicted PD-L1 score compared with others (median range: 0.273 versus 0.185, P= 0.0079). Moreover, patient-level median of predicted PD-L1 scores of lesions was significantly associated with prolonged PFS (hazard ratio 0.69, P= 0.0401) and overall survival (hazard ratio 0.65, P= 0.0431). Conclusions: Deep-learning based imaging biomarker accurately predicts PD-L1 status of each metastasis, as well as independent ICI response reflecting inter-tumoral heterogeneity. This algorithm would guide which lesion would be representative in clinic.