Correlation of fragmented pattern of tumor mass captured by artificial intelligence (AI)-powered whole-slide image (WSI) analysis with biased fibroblast expansion over tumor growth and distinct mutational signatures.

Abstract

e14657 Background: Little is known about the underlying mechanism of how fragmented tumor mass is observed in pathology images. Here, we present the tumor fragmentation index (TFI) defined by the number of separated tumor fragments per total area occupied by tumor, and investigate its association with distinct growth patterns of a variety of cell types in tumor microenvironment (TME) and mutational signature. Methods: Tumor and stromal areas of TCGA H&E whole-slide images (WSI, n = 7472) across the 23 carcinoma cancer types were segmented using Lunit SCOPE, an AI-powered WSI analyzer. Independent tumor masses were isolated by the connected component labeling algorithm. Tumor fragments which are too small in size ( < 968.2 μm2) or which are not in contact with the stromal area were filtered out, then TFI (count / mm2) was calculated. We evaluated the association of the TFI with the proportions of various cell types including tumor cells (TC), fibroblasts (FB), and other cells with nuclei (OT), as well as immune phenotype (IP), and single-base substitution (SBS) mutational signatures. IPs were defined as follows: inflamed IP (IIP) as high intratumoral tumor-infiltrating lymphocyte (iTIL) and stromal TIL (sTIL); immune-excluded IP (IEP) as low iTIL and high sTIL; immune-desert IP (IDP) as low TIL overall. Results: In TCGA pan-carcinoma dataset, median TFI was 12 (interquartile range, IQR 4-34 /mm2), and pancreatic adenocarcinoma (median, [IQR]; 77, [48-112]), prostate adenocarcinoma (58, [28-91]), and breast cancer (57, [22-109]) had top-ranked TFI, whereas renal cell carcinoma (1, [0-3]), and melanoma (1, [0-3]) had the lowest TFI. Considering No. of OT is relatively constant over carcinogenesis process, TFI was negatively correlated with TC-to-OT ratio (rho = -0.338, p < 0.001), but positively correlated with FB-to-OT ratio (rho = 0.312, p < 0.001), reflecting biased fibroblast expansion over tumor growth in the samples with high TFI. TFI was significantly higher in IEP (14, [6-33]) compared to that in the IIP (11, [3-30]) or IDP (9, [1-39], p < 0.001). TFI was positively correlated with SBS30, SBS1, SBS23, and SBS10B, which are dominant for C > T mutation, but negatively correlated with SBS37, SBS26, which are dominant for T > C mutation. Gene set enrichment analysis showed epithelial mesenchymal transition (normalized enrichment score, NES 2.28, adjusted p = 0.019) was up-regulated, but hypoxia (NES -1.49, adjusted p = 0.011) was down-regulated in samples with high TFI (upper 25%). Conclusions: Tumors with highly fragmented patterns in the TME had a distinct patho-genomic profile of faster expansion of fibroblasts, relatively slow-growing tumor cells, and were prone to immune-excluded phenotype. Differential mutational signatures were captured based on TFI score, warranting further investigation.