Prediction of the Occurrence of Microvascular Invasion in Hepatocellular Carcinoma via Preoperative Prediction Model Using Deep Learning on CT Image
10.3969/j.issn.1005-5185.2025.04.009
- VernacularTitle:基于深度学习CT影像术前预测肝癌患者微血管侵犯
- Author:
Xin LI
1
;
Jinming CHEN
;
Rui LIU
;
Le LI
Author Information
1. 内蒙古医科大学赤峰临床医学院,内蒙古 赤峰 024000
- Publication Type:Journal Article
- Keywords:
Carcinoma,hepatocellular;
Microvascular invasion;
Tomography,X-ray computed;
Deep learning;
Forecasting
- From:
Chinese Journal of Medical Imaging
2025;33(4):390-395
- CountryChina
- Language:Chinese
-
Abstract:
Purpose To explore the application value of deep learning in the prediction of hepatocellular carcinoma microvascular invasion based on CT images.Materials and Methods A total of 63 hepatocellular carcinoma patients with pathologic diagnosis proving microvascular invasion status from January 2020 to December 2022 in Chifeng City Hospital were included,and preoperative enhanced CT images of these patients were collected,and randomly divided into training,validation and testing groups.Three classical convolutional neural network models(AlexNet,VGG16 and ResNet50)were established by transfer learning.The general clinical data of hepatocellular carcinoma patients were analyzed by single-factor and multifactor Logistic regression to establish clinical prediction models.The optimal model was obtained by comparing the actual prediction effects of the four prediction models.Results The area under the curve(AUC)of clinical prediction model was 0.844,sensitivity was 0.833 and specificity was 0.708.The AUC of AlexNet model was 0.865,sensitivity was 0.833,specificity was 0.717 and training time was 818.2 s.The AUC of VGG16 model was 0.892,sensitivity was 0.857,specificity was 0.717 and training time was 9 743.2 s.The AUC of ResNet50 model was 0.937,sensitivity was 0.786,specificity was 0.925 and training time was 3 800.8 s.Conclusion Deep learning modeling is an assessment method for preoperative noninvasive prediction of microvascular invasion,which is superior to clinical prediction models,with the best performance of the ResNet50 model.It can provide a strong reference for physicians in making clinical decisions.
