Abstract
BACKGROUND: Computer-aided histopathological image analysis is increasingly used for image evaluation and decision-making in cancer patients. This study extracted quantitative histopathological image features to predict molecular features, and combined them with omics data to predict prognosis of hepatocellular carcinoma (HCC) patients. METHODS: Totally 334 patients from The Cancer Genome Atlas were divided equally into the training and testing sets. Histopathological image features and multiple omics data (somatic mutation, mRNA expression, and protein expression) were used alone or in combination to build prediction models through machine learning. Areas under receiver operating characteristic curves (AUCs) were assessed for 1-year, 3-year, and 5-year overall survival (OS). RESULTS: Histopathological image features were able to predict somatic mutations: TERT promoter (AUC = 0.926), TP53 (AUC = 0.893), CTNNB1 (AUC = 0.885), ALB (AUC = 0.879), molecular subtypes (AUCs from 0.905 to 0.932), and OS (5-year AUC = 0.819) in the testing set, which also had good performances for OS in the external validation sets of tissue microarrays from 263 patients (5-year AUCs from 0.682 to 0.761). Furthermore, the integrated models of histopathological image features and omics data increased the accuracy of prognosis prediction, especially the multi-platform model that combined all types of features (5-year AUC = 0.904). The risk score based on the multi-platform model was a significant predictor for OS in the testing set (HR = 15.09, p < 0.0001). Additionally, the multi-platform model achieved a higher net benefit in decision curve analysis. CONCLUSION: Histopathological image features had the potential to predict molecular features and survival outcomes, and could be integrated with multiple omics data as a practical tool for prognosis prediction and risk stratification, facilitating personalized medicine for HCC patients.