Abstract 453: Establishing quantitative image analysis methods for tumor microenvironment evaluation

Abstract

Abstract Introduction: There is growing evidence that supports the role of the tumor microenvironment (TME) in the development and progression of hepatocellular carcinoma (HCC). However, the correlation between its composition and prognosis remain unclear. TME evaluation requires a combination of cell type and spatial information. These information can be obtained with the use of immunohistochemistry on patient derived slides. However, the IHC quantification remains a challenge. Computational methods such as artificial intelligence-based tool, may expedite the detection and classification thousands of different cells, expanding our understanding of the TME. Here, we aim to develop an AI based image analysis pipeline to define morphological and immunological characteristics of HCC-TME, as well as their relationship with clinicopathological features. Materials and Methods: We collected 98 HCC samples from liver resections available in the . TME composition of the tumors was evaluated and classified as inflamed, immune excluded and immune desert. Tumor slides were stained with a panel of TME markers (CD3, CD8, FOXP3, TIGIT, RORgt, ICOS, GranzymeB CD163, iNOS, PD-1, PD-L1) by IHC. The slides were digitalised and whole slide images were used for the quantification. The samples were split into training (80%) and test (20%) datasets and used to train convolutional neural network (CNN) models. For the quantification of immune cells, we trained two separate CNNs: cell detection and tumor-stroma segmentation. Cell nucleus instance segmentation was achieved using StarDist package (Schmidt et al 2018). We trained a model with our slides and tested the pretrained model (2D_versatile_he) which is for H&E stained images. Immune cells were classified using random forest classifier in QuPath. Finally, we trained a CNN in UNET architecture with ResNet34 backbone for semantic segmentation of tumor tissue into parenchyma, stroma and debris classes, by fastai deep learning library. Results: The accuracy of pretrained StarDist model was limited to 72% on IHC slide images. Thus, we trained a new cell nucleus instance segmentation StarDist model with our dataset and it reached 84% accuracy, 91.3% F1-Score, 92% true positive, 90.6% true negative rates on IHC slide images. Random forest classifier annotated immune cells at 98% accuracy. The tissue segmentation model classified tumor regions into parenchyma, stroma and debris at 95,8% accuracy, 92.5% dice, 86.3% IuO. Conclusions: In this study we developed a pipeline implementing open-source solutions to quantify IHC slides. The use of this semi-automatized computational pathology workflow can provide robust information in regard of the TME composition augmenting the discovering tumour specific TME features and pave way for the discovery of novel prognostic and therapeutic targets. Citation Format: Caner Ercan, Mairene Coto-Llerena, Salvatore Piscuoglio, Luigi M. Terracciano. Establishing quantitative image analysis methods for tumor microenvironment evaluation [abstract]. In: Proceedings of the American Association for Cancer Research Annual Meeting 2022; 2022 Apr 8-13. Philadelphia (PA): AACR; Cancer Res 2022;82(12_Suppl):Abstract nr 453.