Abstract PR-03: Radiomics and AI-based treatment decision support for non-small cell lung cancer

Abstract

Abstract Two treatment strategies, tyrosine kinase inhibitors (TKIs), and immune checkpoint inhibitors (ICIs), have improved long-term survival for patients with advanced non-small cell lung cancer (NSCLC). However, choosing the right treatment is challenging as the currently approved biomarkers for treatment decision are biopsy based: activating mutations in epidermal growth factor receptor (EGFR) and expression of programmed death-ligand 1 (PD-L1). Biopsies are subject to false negatives because of sampling bias, and they can change during therapy rendering that treatment ineffective. Therefore, there is a compelling need to identify additional biomarkers that reflect the entire tumor and can be obtained non-invasively to help guide therapy choice. Herein, we test the hypothesis that EGFR mutation and PD-L1 expression status can be captured by analysis of 18F-FDG-PET/CT images with artificial intelligence (AI) algorithms. These can be used in combination to construct a treatment decision support system based on their association with clinical outcomes following treatment with TKIs or ICIs. This retrospective study enrolled 837 NSCLC patients from four international institutions: Shanghai Pulmonary Hospital (SPH), the Fourth Hospital of Hebei Medical University (HBMU), the Fourth Hospital of Harbin Medical University (HMU), and the H. Lee Moffitt Cancer Center (HLM). Two residual-convolutional-network (ResCNN) models to predict EGFR mutation and PD-L1 positive status were trained (N=429) and validated (N=187) with PET/CT images and clinical data of 616 patients from SPH and HBMU, and then tested using external HMU test cohort with EGFR mutation status (N=65) and HLM test cohort with PD-L1 expression status (N=85). Subsequently, the generated EGFR and PD-L1 deep learning scores (EGFR-DLS and PDL1-DLS) of the ResCNN models were further associated with progression-free survival (PFS) in HMU EGFR-TKI-treated patients (N=67) and HLM ICI-treated patients (N=149). The EGFR-DLS and PDL1-DLS demonstrated high accuracy in EGFR mutation status prediction with AUCs of 0.86 (95%CI: 0.83-0.90), 0.83 (95%:0.78-0.88) and 0.81(95%CI: 0.72-0.91), and in PD-L1 positive status discrimination with AUCs of 0.89 (95%CI: 0.84-0.94), 0.84 (95%CI: 0.76-0.92) and 0.82 (95%CI: 0.74-0.89) for training, validation and external test cohorts, respectively. Patient responses to TKI and ICI were found to be reciprocal: for patients with high EGFR-DLS, EGFR-TKI treatment was significantly associated with longer PFS regardless of PDL1-DLS (low PDL1-DLS, p=0.006; high PDL1-DLS, p=0.067); for patients with low EGFR-DLS, ICI treatment was significantly associated with longer PFS (p=0.007) in particular those patients with high PDL1-DLS. Therefore, the combination of EGFR-DLS and PDL1-DLS could be used as an alternative non-invasive Decision Support Tool for NSCLC. Citation Format: Wei Mu, Lei Jiang, Jianyuan Zhang, Yu Shi, Jhanelle E. Gray, Xinming Zhao, Xilin Sun, Jie Tian, Matthew B. Schabath, Robert J. Gillies. Radiomics and AI-based treatment decision support for non-small cell lung cancer [abstract]. In: Proceedings of the AACR Virtual Special Conference on Artificial Intelligence, Diagnosis, and Imaging; 2021 Jan 13-14. Philadelphia (PA): AACR; Clin Cancer Res 2021;27(5_Suppl):Abstract nr PR-03.