Abstract 1394P
Background
Immune checkpoint inhibitors (ICI) represent a major change in non-small cell lung cancer (NSCLC) treatment, however robust biomarkers are needed. Emerging data suggest that features discovered by deep learning (DL) models from CT scan images using artificial intelligence (AI) algorithms can accurately predict outcomes. In this study, our objective was to explore the potential of AI-based DL radiomics models in patients with advanced NSCLC treated with ICI.
Methods
Pre-ICI CT scans were obtained from n=141 patients with advanced NSCLC. Primary lung cancer lesions were annotated using Weasis medical viewer. PyRadiomics analyses were performed using published known open-source algorithms. In parallel, DeepRadiomics based on convolutional AI neural networks using the Imagia Evidens platform were used. Algorithms were designed by incorporating the baseline clinical characteristics and DL-features in multivariate models to predict overall-response rate (ORR), progression-free survival (PFS) and overall survival (OS).
Results
Clinical characteristics (area under the receiver operating characteristic curve [AUC] 0.63, 95%CI 0.53-0.73) or PyRadiomics (AUC 0.67, 95%CI 0.57-0.77) alone were moderately predictive of ORR whereas DeepRadiomics, combined with clinical characteristics, had an AUC of 0.78 (95%CI 0.70-0.85). At a median follow-up of 50 weeks, the association between PFS and clinical characteristics was comparable to that of DeepRadiomics. Further, PFS prediction was not improved when clinical characteristics were combined with DeepRadiomics. With respect to OS, seven DeepRadiomics features were associated with increased hazard of death (HR 1.72-11.50; all p<0.05) while 6 were associated with a decreased hazard (HR 0.08-0.52); p<0.05). Of these 13 features, 11 maintained their association with OS after adjustment with clinical characteristics.
Conclusions
These preliminary results reveal the potential of deep learning methods to improve upon standard predictive factors for ICI response. Using DeepRadiomics features to predict clinical outcomes may represent a novel clinical approach in the immuno-oncology arena.
Clinical trial identification
Editorial acknowledgement
Legal entity responsible for the study
Centre de recherche de l'Université de Montréal.
Funding
Centre de recherche de l'Université de Montréal.
Disclosure
A. Elkrief: Research grant/Funding (institution), Not related to submitted work.: AstraZeneca. K. Phan, L. Di Jorio, R. Simpson, M. Kosyakov, F. Chandelier, K. Kafi: Full/Part-time employment: Imagia. B. Routy: Research grant/Funding (institution): Imagia. All other authors have declared no conflicts of interest.