Objective: Stromal elements in the tumor microenvironment (TME) impact prognosis and response to therapy, creating a barrier to absorption and penetration of therapeutic drugs and modulating immune cell infiltration. Few models exist to analyze the spatial distribution and mapping of stroma and cancer lesions composing the complex TME in association with mRNA levels via high throughput deep learning algorithm. Methods: The histopathology images (H&E stain) and mRNA-seq of 625 CRC patients (pts) were obtained from the Cancer Genome Atlas (TCGA). A deep learning-based model of tissue classification enabling segmentation of tumor (malignant glands composed of adenocarcinoma cells) and stroma in histopathology images was established using published data sets [1-2]. We defined the following parameters based on the pixel count of stroma and tumor: the stroma-to-tumor ratio (STR) defined as the pixel count ratio of stroma to stroma plus tumor; STRx, STR within the pixel distance x from the tumor; tumor-tissue-ratio (TTR), a pixel count ratio of tumor to stroma plus tumor. The expressivity of genes enriched in cellular and structural components of TME and its correlation with STR and TTR were analyzed. Results: The profiling results demonstrate that histopathology images of consensus molecular subtype (CMS) 4 of colorectal cancer have statistically significantly larger STR, STR10, STR20, STR30, STR40, and STR50 than those in other CMS groups (p<0.0001); TTR, smaller in CMS 4 than in other CMS groups (p<0.0001). Out of two patient groups clustered via k-means clustering (high and low STR), the group having higher STR has significantly decreased expression of genes related to antigen presentation, natural killer cells, and plasmacytoid dendritic cells, as well as significantly elevated expression of genes associated with epithelial-mesenchymal transition, desmoplastic reaction, fibroblastic cytokines, and angiogenesis (p<0.05) (refer to presentation, regarding the list of genes).

Conclusion: Our analysis of histopathology images via the deep learning-based algorithm enables automated segmentation of stroma and malignant lesions. Parameters representing spatial and architectural distribution of stroma and cancer lesions are shown to have close correlation with varying expression of genes associated with immune system activation and suppression in the tumor microenvironment.

[1] Kather et al. Predicting survival from colorectal cancer histology slides using deep learning: A retrospective multicenter study. PLoS Medicine. 2019

[2] Kumar et al. A Multi-organ nucleus segmentation challenge. IEEE Trans Med Imaging. 2019

Citation Format: Yeongwon Kim, Kyungdoc Kim, Jeonghyuk Park, Hyunho Park, Kyu-Hwan Jung, Sunyoung S. Lee. Deep learning-based analysis of tissue segmentation in histopathology images of colorectal cancer [abstract]. In: Proceedings of the Annual Meeting of the American Association for Cancer Research 2020; 2020 Apr 27-28 and Jun 22-24. Philadelphia (PA): AACR; Cancer Res 2020;80(16 Suppl):Abstract nr 2631.