1219P - Integrating artificial intelligence (AI)-based lymphocytic infiltration assessment in early stage NSCLC: A sub-study of the TNM-I trial

Background

Tumor infiltrating lymphocytes (TILs) serve as both prognostic and predictive biomarkers for immunotherapy responses in various cancers, including non-small cell lung cancer (NSCLC). The TNM-I study (NCT03299478), a large multi-institutional Scandinavian trial, aims to develop an immune-cell scoring system for operable NSCLC. To overcome the manual limitations in TILs assessment, a well-validated and benchmarked deep learning (DL) pipeline has been developed.

Methods

We introduced a fully automated AI model named Deep-ICP to predict TILs from whole-slide images of H&E-stained tissues. This involves tissue segmentation, cell detection and classification. The model was trained and tested > 2 million cells (including TILs, tumor cells, and stromal cells) as ground truth from 446 NSCLC patients at Dana Farber Cancer (DFCI) Institute. The model was further validated in a TNM-I prospective cohort of 724 patients. Deep-ICP outputs were benchmarked against highly-cited DL algorithms and validated using NanoString gene expression immune signatures and multiplex IHC for CD8+pCK.

Results

Deep-ICP outperformed existing models like HoverNet and Stardist in cell detection (Dice score: 0.89, AJI: 0.72, PQ: 0.74) and classification (F1 score: 0.88, accuracy: 0.83) on the DFCI training set. In the TNM-I cohort, the median number of TILs per mm² was 822 (IQR: 540–1247). A moderate correlation was found between TIL density and CD8+ cell/mm² density (r=0.41), and between the percentage of Deep-ICP-detected tumor cells and pCK+ tumor cells (r=0.44). In a subset of the TNM-I cohort using gene expression analysis (n=122), moderate correlations were noted between TILs and adaptive immunity gene signatures including CD45-, exhausted CD8-, and T- & B- cells (r=0.37 to 0.54; p.adjust <0.0001). No significant associations were observed with innate immunity signatures such as neutrophils, macrophages, dendritic cells, and mast cells.

Conclusions

The Deep-ICP AI model demonstrated superior performance in histology-based cell classification compared to widely used pipelines. This innovative approach provides a robust framework for integrating AI into the TNM-I immune cell scoring system for NSCLC.

Clinical trial identification

NCT03299478.

Editorial acknowledgement

Legal entity responsible for the study

The authors.

Funding

Norwegian Cancer Society, Helse Nord RHF.

Disclosure

A. Helland: Financial Interests, Institutional, Advisory Board, Advisory boards: Janssen, Takeda, AstraZeneca, AbbVie, Roche, BMS, Pfizer, MSD, Bayer, Lilly, Medicover; Financial Interests, Institutional, Invited Speaker, talks at meetings: AstraZeneca, Roche, AbbVie, Pfizer; Financial Interests, Institutional, Coordinating PI, BMS provides drug to patients in an investigator initiated clinical trial: BMS; Financial Interests, Institutional, Coordinating PI, Ultimovacs provides drug and funds for investigator initiated clinical trial: Ultimovacs; Financial Interests, Institutional, Coordinating PI, AstraZeneca provides drug and funds for investigator initiated clinical trial: AstraZeneca; Financial Interests, Institutional, Coordinating PI, Roche provides drug and funds for investigator initiated clinical trial: Roche; Financial Interests, Institutional, Coordinating PI, Novartis provides drug and funds for clinical trial: Novartis; Financial Interests, Institutional, Coordinating PI, Eli Lilly provides drug and funds for clinical study: Eli Lilly; Financial Interests, Institutional, Coordinating PI, Incyte provides drug and funds for clinical study: Incyte; Financial Interests, Institutional, Coordinating PI, Illumina provides assays for patients in a clinical trial: Illumina; Financial Interests, Institutional, Coordinating PI, GSK provides drug and funds for investigator initiated clinical trial: GSK; Non-Financial Interests, Other, Board member in the patient organisation until 2022. Provides advice and gives talks.: The lung cancer patients organisation. D. Kwiatkowski: Non-Financial Interests, Institutional, Research Funding: Genentech, AADI, Revolution Medicines; Non-Financial Interests, Institutional, Advisory Role: Genentech, AADI, Expertconnect, Guidepoint, Bridgebio, Slingshot Insights, William Blair, MEDACorp, Radyus Research. All other authors have declared no conflicts of interest.