Leveraging deep learning to discover interpretable cellular spatial biomarkers for prognostic predictions based on hepatocellular carcinoma histology

Abstract

The spatial structure of various cell types in the tumour microenvironment (TME) can provide valuable insights into disease progression. However, identifying the spatial organization of diverse cell types that significantly correlates with patient prognosis remains challenging. In this study, enabled by deep learning-based cell segmentation and recognition, we developed a computational pipeline to systematically quantify the spatial distribution features of tumour cells, stromal cells, and lymphocytes in haematoxylin and eosin (H&E)-stained pathological images of hepatocellular carcinoma (HCC). We identified six cellular spatial features that consistently and significantly correlated with the overall survival of patients in two independent HCC patient cohorts, The Cancer Genome Atlas Program cohort and the Beijing Hospital cohort. Each threshold for patient stratification was the same for both cohorts, and the six features independently served as prognostic indicators when individually analysed alongside clinical variables. Furthermore, the combination of features such as the mean value of cellular diversity around stromal cells (StrDiv-M), the median distance between all cells (CellDis-MED), and the median value of variation coefficient of the distance around stromal cells and their neighbours (CvStrDis-MED) could further stratify the patient prognosis. In addition, incorporating cell spatial features with another clinical feature, microvascular invasion improved prognostic stratification efficacy for patients from both cohorts. In conclusion, by quantifying the cellular spatial organization features in the HCC TME, we discovered novel biomarkers for evaluating tumour prognosis. These findings could promote mechanistic studies of the cellular spatial organization within the HCC TME and potentially guide future clinical treatment.