Abstract 4170: Artificial intelligence (AI)-based multi-modal approach using H&E and CT image for predicting treatment response of immune checkpoint inhibitor (ICI) in non-small cell lung cancer (NSCLC)

Abstract

Abstract Background: Previously, AI-powered spatial tumor-infiltrating lymphocyte analysis in H&E whole-slide image (WSI) showed promising performance in predicting ICI outcomes. However, there were several innate limitations of tissue analysis, including the time lag between tissue biopsy and actual ICI administration, during which tumor microenvironment (TME) has changed, and analyzing a part of a single tumor tissue would not fully assess immunogenicity of multiple metastatic tumors. To overcome the challenges, we developed multiple-instance learning (MIL)-based models and applied them to predict the treatment response of ICI in a real-world dataset. Method: MIL-based models were trained under the cross-validation scheme for CT scans (N=456) separately, based on CT image analyzers. Lunit SCOPE IO (Lunit, South Korea), an AI-based H&E WSI analyzer, identifies various classes of cells within TME and segments tumor regions from WSI (AI-HE-TIL). An AI-based CT image analyzer detects nodules on the CT image (AI-CT). The final ensemble score per patient is computed by the average of all model predictions from CT and pathology image inputs. Single and ensemble performances for predicting responder and non-responder patients to ICI are measured in an independent cohort of 208 NSCLC patients. Results: In the overall cohort, the AUROC values for AI-HE-TIL, AI-CT, and the ensemble were 0.6372, 0.6759, and 0.7307, respectively. The high ensemble score group (≥ median) has prolonged progression-free survival (PFS) of ICI (median PFS 5.0 months vs 2.1 months, hazard ratio [HR] 0.54, P < 0.0001). It correlates with PD-L1 expression, as tumor proportion score (TPS) ≥ 50% subgroup had higher ensemble score (median 0.312 vs 0.257, P = 0.0295). Interestingly, among TPS ≥ 50% subgroup, the high ensemble score shows significantly favorable PFS of ICI (median PFS 6.7 months vs 2.9 months, HR 0.46, P = 0.0002). In subgroup analysis according to tissue harvest site (primary tumor vs lymph node or distant metastasis), the AUROC values of AI-HE-TIL and AI-CT for the subgroup harvested from the primary tumor (N=111) were 0.6568 and 0.6658, respectively, and those for the subgroup harvested from the metastatic lesion (N=97) were 0.6162 and 0.6973, respectively, implying AI-CT is relatively not affected by tissue harvest site. Conclusion: AI-powered multi-modal approach of using H&E image and CT image shows complimentary, and synergistic effect to predict ICI clinical outcomes in advanced NSCLC, even in the subgroup of PD-L1 high population. Citation Format: Dong Young Jeong, Jongchan Park, Heon Song, Jimin Moon, Taebum Lee, Changho Ahn, Sehhoon Park, Se-Hoon Lee, Chan-Young Ock, Ho Yun Lee. Artificial intelligence (AI)-based multi-modal approach using H&E and CT image for predicting treatment response of immune checkpoint inhibitor (ICI) in non-small cell lung cancer (NSCLC) [abstract]. In: Proceedings of the American Association for Cancer Research Annual Meeting 2024; Part 1 (Regular Abstracts); 2024 Apr 5-10; San Diego, CA. Philadelphia (PA): AACR; Cancer Res 2024;84(6_Suppl):Abstract nr 4170.