Abstract 1357P
Background
Recent studies suggest imaging-based signatures can be a promising non-invasive approach to predict response to immune checkpoint inhibitors (ICI). In this multi-centric study, we explored the use of radiomics models to characterize the relationship between PD-L1 expression, imaging features, and clinical outcomes.
Methods
Primary lung lesion on pre-ICI CT scans were semi-automatically annotated from 299 advanced NSCLC patients from two institutions (n=171 CHUM and n=128 JGH). Multivariate models incorporating clinical variables (e.g., ECOG, 1st line vs 2nd line, stage, smoking status), PD-L1 expression, and PyRadiomic (PR) or DeepRadiomics (DR) features were developed in a discovery cohort (80%) to predict ORR, PFS at 6 months, OS at 1 year and were evaluated in a validation cohort (20%).
Results
Baseline characteristic and clinical outcomes were well balanced, moreover, principal component analysis did not reveal difference in radiomics features between both centers. The ORR was 55.1% and the median PFS and OS were 5.4 and 14.3 months, respectively. DR features were significantly superior to predict PD-L1 ≥ 50% compared to PR in the discovery cohort (AUC 0.81 [95%CI 0.76-0.87] vs 0.72 [95%CI 0.64-078]). Their performance in the validation cohort were 0.63 [95%CI 0.51-0.79] and 0.61 [95%CI 0.47-0.75], respectively. Incorporation of radiomic features improved the discrimination of ORR, PFS, and OS beyond clinical variables and PD-L1 prediction potential in the discovery cohort. For ORR and PFS neither PR nor DR improved prediction on validation cohort. However, in the validation cohort the AUC for OS of clinical variables and PD-L1 was AUC 0.68 (95%CI 0.54-0.80) compared to 0.59 (95%CI 0.46-0.73) and 0.66 (95%CI 0.51-0.83) with PR and DR, respectively.
Conclusions
These results show the added potential of deep radiomics methods in predicting PD-L1 expression as well as ICI response in addition to clinical variables while demonstrating the challenges of their generalizability. Expansion of the study is underway to improve sample size and explore novel feature engineering methods better suited for generalization.
Clinical trial identification
Editorial acknowledgement
Legal entity responsible for the study
B. Routy.
Funding
FRQS - Concours Onco-Tech.
Disclosure
K. Phan: Other, Personal, Full or part-time Employment: Imagia. L. Ding: Financial Interests, Personal, Full or part-time Employment: Imagia. T. Nair: Financial Interests, Personal, Full or part-time Employment: Imagia. F. Chandelier: Financial Interests, Personal, Full or part-time Employment: Imagia. K. Kafi: Financial Interests, Personal, Full or part-time Employment: Imagia. L. Di Jorio: Financial Interests, Personal, Full or part-time Employment: Imagia. All other authors have declared no conflicts of interest.