A prediction model based on contrast-enhanced MRI radiomics and clinical features for early recurrence of hepatocellular carcinoma after radical resection
10.3760/cma.j.cn113884-20220513-00210
- VernacularTitle:基于增强MRI影像组学及临床特征构建肝细胞癌根治术后早期复发预测模型的价值
- Author:
Yang GAO
1
;
Chuanqiang LAN
;
Weichuan YE
;
Yumin HU
;
Jianjian XING
;
Yongjin ZHOU
;
Jingle FEI
;
Jiansong JI
Author Information
1. 温州医科大学附属第五医院放射科 浙江省影像诊断与介入微创研究重点实验室，丽水 323000
- Keywords:
Carcinoma, hepatocellular;
Magnetic resonance imaging;
Radiomics;
Early recurrence
- From:
Chinese Journal of Hepatobiliary Surgery
2022;28(11):817-821
- CountryChina
- Language:Chinese
-
Abstract:
Objective:To develop a prediction model based on imaging features by contrast-enhanced MRI radiomics combined with clinical features for early recurrence of hepatocellular carcinoma (HCC) after radical resection.Methods:A retrospective study was carried out on 109 HCC patients who underwent radical resection at the Fifth Affiliated Hospital of Wenzhou Medical University from January 2015 to December 2020. Of 109 patients enrolled in this study, there were 96 males and 13 females, aged (58.3±10.7) years. Based on whether there was recurrence within 12 months after operation, the patients were divided into the early recurrence group ( n=31) and the control group ( n=78). These 109 patients were then randomly divided into the validation set ( n=23) and the training set ( n=86) at a ratio of 1∶4. Based on preoperative multi-phase contrast-enhanced MRI scanning, the tumor lesions were delineated on the Radcloud platform, and 1 409 quantitative radiomic features were extracted. Dimension reduction and screening of these features were carried out using variance threshold, SelectKBest and LASSO. Combined with clinical features (alpha fetoprotein, tumor size), several prediction model were established through machine learning. The predictive efficiencies of these models were evaluated using the area under the receiver operating characteristic (ROC) curve, accuracy rate, recall rate and balanced F score. Results:The proportions of irregular tumor shape and unclear tumor boundary, as well as maximum tumor diameter in the early recurrence group were significantly higher than that in the control group, but the proportion of pseudocapsule was significantly lower than that in the control group (all P<0.05). A total of 465 features were screened from the 1 409 features using the variance threshold method, followed by 38 features were screened using the method of SelectKBest. Finally 7 optimal radiomic features were screened based on the LASSO method. When combined with clinical features, 5 prediction models were established through machine learning. These models were support vector machine, Gaussian naive bayes, logistic regression, Multinomial naive bayes and K-nearest neighbor (KNN), respectively. Among these 5 models, the prediction efficiency of the KNN model was relatively highest, with the area under the ROC curve, accuracy rate, recall rate and balanced F score being 0.90, 0.98, 0.74 and 0.84 in the training set, and 0.76, 0.92, 0.75 and 0.83 in the verification set, respectively. Thus, the KNN model was selected as the best prediction model in this study. Conclusion:The prediction model of KNN was developed for early recurrence of HCC after radical resection based on preoperative contrast-enhanced MRI radiomics combined with clinical features.
