Prediction of immunotherapy response using deep learning of PET/CT images

Currently only a fraction of patients with non-small cell lung cancer (NSCLC) experience durable clinical benefit (DCB) from immunotherapy, robust biomarkers to predict response prior to initiation of therapy are an emerging clinical need. PD-L1 expression status from immunohistochemistry is the only clinically approved biomarker, but a non-invasive complimentary approach that could be used when tissues are not available or when the IHC fails and can be assessed longitudinally would have important implications for clinical decision support. In this study, 18F-FDG-PET/CT images and clinical data were curated from 697 NSCLC patients from three institutions. Utilizing PET/CT images, a deeply-learned-score (DLS) was developed by training a small-residual-convolutional-network model to predict the PD-L1 expression status, which was further used to predict DCB, progression-free survival (PFS), and overall survival (OS) in both retrospective and prospective test cohorts of immunotherapy-treated patients with advanced stage NSCLC. This PD-L1 DLS significantly discriminated PD-L1 positive and negative patients (AUC[≥]0.82 in all cohorts). Further, higher PD-L1 DLS was significantly associated with higher probability of DCB, longer PFS, and longer OS. The DLS combined with clinical characteristics achieved C-indices of 0.86, 0.83 and 0.81 for DCB prediction, 0.73, 0.72 and 0.70 for PFS prediction, and 0.78, 0.72 and 0.75 for OS prediction in the retrospective, prospective and external cohorts, respectively. The DLS provides a non-invasive and promising approach to predict PD-L1 expression and to infer clinical outcomes for immunotherapy-treated NSCLC patients. Additionally, the multivariable models have the potential to guide individual pre-therapy decisions pending in larger prospective trials.