Exploratory analysis in quantitative radiomics in hepatocellular carcinoma

Abstract

Objective: Hepatocellular carcinoma (HCC) is one of the common cancers in the world with high percentage of post-surgical recurrence rate. The presence of microvascular invasion (MVI) is known to be a poor prognostic feature and is traditionally confirmed by histopathological tests of the resected specimens or invasive liver biopsy which may potentially pose the risk of tumor seeding. To have optimised patient specific treatment outcome and relieve the stress of organ shortage, it is desirable to determine the presence of MVI noninvasively at pre-operational stage. In the present study, exploratory analysis of quantitative radiomics of magnetic resonance imaging (MRI) was introduced as anon-invasive and pre-operative method for the prediction of MVI in HCC. The goal of this study is to discern for the predictive ability of proposed models using supervised machine learning approach for prediction of MVI in HCC. The result may potentially act as preliminary information for large scale study to establish a robust model to predict prognosis and identify patient in whom curative resection and transplantation is more desirable treatment of HCC, thus allow treatment optimisation and relieve organ shortage. Method and Materials: 14patients with HCC who performed surgical resection or liver transplantation were included in the study. Liver images acquired pre-operatively on a 3T whole body MRI scanner were retrieved retrospectively. Regions of interest (ROIs) were selected on the MR images where several feature extraction methods were applied. Classifications concerning the textural features by supervised learning approaches followed by those with principal component analysis (PCA) were stimulated. Result: Four classifiers, which were fine gaussian SVM, medium kNN, coarse kNN and ensemble boosted trees, were found to maintain a stable accuracy of 86.7%irrespectiveof the means of feature extraction methods. With the incorporation of principal component analysis (PCA), the average accuracy increased from 76.3% to 82%. Conclusion: Quantitative texture analysis can be considered as a potential reproducible method for prediction of MVI in HCC for decision support to optimize treatment outcome. Four models with classifiers of gaussian SVM, medium kNN, coarse kNN or ensemble boosted trees were found to be robust models with high accuracy of 86.7% irrespective of feature extraction methods for prediction of MVI in HCC.