Home > Published Issues > 2019 > Volume 7, No. 4, December 2019 >

Comparison of Machine Learning-Based Radiomics Models for Early Recurrence Prediction of Hepatocellular Carcinoma

Panyanat Aonpong 1, Qingqin Chen 2, Yutaro Iwamoto 1, Lanfen Lin 3, Hongjie Hu 2,Qiaowei Zhang 2,
and Yen-Wei Chen 1,4,3
1. College of Information Science and Engineering, Ritsumeikan University, Shiga, Japan
2. Sir Run Run Shaw Hospital, Zhejiang University, Hangzhou, China
3. College of Computer Science and Technology, Zhejiang University, Hangzhou, China
4. Zhejiang Lab, Hangzhou, China

Abstract—BACKGROUND & AIMS: Using a radiomics model, we investigated computed tomography images to make a preoperative prediction of the Early Recurrence (ER) of Hepatocellular Carcinoma (HCC). A radiomics model mainly consists of feature extraction, feature selection, and classification. The conventional method used Least Absolute Shrinkage and Selection Operator (LASSO) regression to select the features and the classification. METHODS: We compared the new combination of feature selection and classification methods for the preoperative ER prediction of HCC. The new combination gave a significantly higher accuracy than the conventional method. Twelve combination models were provided by using different combinations of the feature selection methods and the classification methods. We used three classification methods, which included LASSO, linear Support Vector Machines (SVMs), and decision trees. We compared the performance of each method by using the area under the curve of the receiver operating characteristic to show a more appropriate way to detect the ER of HCC. In addition, we compared the efficiency term of each feature elimination method. These two comparisons can measure the quality of feature selection and the compatibility between both the feature selection and classification phases. RESULTS: Approximately 65 features were selected from 300 features. Our proposed combination showed that the accuracy could be improved by using the SVM classification method for new radiomics models. The accuracy of feature selection using LASSO regression with Support Vector Machine classification can reach 0.8918, whereas this accuracy reaches 0.8779 when support vector machine-recursive feature elimination is used with LASSO classification.

Index Terms—hepatocellular carcinoma, early recurrence, preoperative prediction, liver cancers, Radiomics, feature extraction, feature selection, feature elimination, classification, machine learning

Cite: Panyanat Aonpong, Qingqin Chen, Yutaro Iwamoto, Lanfen Lin, Hongjie Hu, Qiaowei Zhang, and Yen-Wei Chen, "Comparison of Machine Learning-Based Radiomics Models for Early Recurrence Prediction of Hepatocellular Carcinoma," Journal of Image and Graphics, Vol. 7, No. 4, pp. 117-125, December 2019. doi: 10.18178/joig.7.4.117-125