Prediction of hepatocellular carcinoma and Edmondson-Steiner grade using an integrated workflow of multiple machine learning algorithms

Abstract

Abstract Early diagnosis of hepatocellular carcinoma (HCC) is indeed a great challenge. Based on traditional methods, the specificity and sensitivity of US/AFP are insufficient to detect the early onset of HCC. In this study, we constructed a prediction model for HCC diagnosis and Edmondson-Steiner (ES) grade using machine learning algorithms. The prediction model was constructed based on CT/MRI images, blood AFP, and pathological diagnosis datasets of 171 patients from Zhejiang Provincial People's Hospital. First, the automatic liver segmentation method of deep learning algorithm is used to locate the region of interest, and then PyRadiomics (engineering hard-coded feature algorithm) and Boruta (random forest algorithm) are used to extract and screen disease-related image features. By comparing the performance of various algorithms, we choose "plr" as the optimal algorithm for the HCC diagnosis model with AUC of 0.990, Kappa of 0.893 and accuracy of 0.952. "gbm" is the optimal algorithm for the ES grade prediction model with AUC 0.941, Kappa 0.777, and accuracy rate 0.902 in the TCGA-LIHC dataset. Compared with traditional diagnostic models based on clinical features, our model significantly improves the predictive performance. AUC increased from 0.733 to 0.933. This study shows that processing image data using deep learning methods can yield important features compared to conventional methods. Choosing an appropriate machine learning algorithm to build a predictive model can significantly improve the performance of disease diagnosis.