Cell graph analysis in hepatocellular carcinoma: predicting local recurrence and identifying spatial relationship biomarkers.

A whole pathology section contains approximately 1,000,000 cells of various types, this large-scale heterogeneity of cells and non-cellular constituents constructs a mutually competitive community. Conventional pixel-based visual processing techniques are insufficient to accurately capture the complexities inherent with cell-entity deployment and formation strategy. Here, we conquered segmentation and classification of all cells on the whole pathology sections from 387 hepatocellular carcinoma (HCC) patients across six cohorts with 57 pathologists assisted. Further, an AI system called Hybrid Graph Neural Network-Transformer system (HGTs) was proposed. It precisely predicted local recurrence of postoperative HCC by analyzing cell interactions across multiple scales, from cell-to-cell, cell-community, to tissue-level interactions. The proposed HGTs outperformed existing SOTA methods, with the C-index improving by 23.1% to reach 0.823, by further integrating multimodal data, including clinical information and immunohistochemical markers. A set of spatial relational biomarkers influencing tumor prognosis was discovered and quantitatively validated. They include the frequency of tumor-lymphocyte and tumor-tumor interactions, the distribution and sparsity of key cellular communities, and the degree of fibrosis in adjacent peritumoral tissues. Utilizing the anti-tumor potential of this marker set, we're developing therapies to enhance the immune system's fight against cancer. All cell semantic segmentation datasets and code are publicly available: https://github.com/Yuan1z0825/HGTs .