Radiomics and deep transfer learning based on gadoxetic acid disodium-enhanced MRI for predicting preoperative microvascular invasion in hepatocellular carcinoma(PDF)
《中国医学物理学杂志》[ISSN:1005-202X/CN:44-1351/R]
- Issue:
- 2025年第10期
- Page:
- 1353-1360
- Research Field:
- 医学影像物理
- Publishing date:
Info
- Title:
- Radiomics and deep transfer learning based on gadoxetic acid disodium-enhanced MRI for predicting preoperative microvascular invasion in hepatocellular carcinoma
- Author(s):
- CHEN Zhao; ZHANG Yu; ZHOU Le; CHEN Qiang; SU Huawei
- Department of Radiology, the Affiliated Hospital of Qingdao University, Qingdao 266555, China
- Keywords:
- Keywords: hepatocellular carcinoma microvascular invasion radiomics deep transfer learning gadoxetic acid disodium
- PACS:
- R318；R816.5
- DOI:
- DOI:10.3969/j.issn.1005-202X.2025.10.013
- Abstract:
- Abstract: Objective To explore the value of radiomics and deep transfer learning (DTL) based on gadoxetic acid disodium-enhanced magnetic resonance imaging (MRI) for preoperative prediction of microvascular invasion (MVI) in hepatocellular carcinoma (HCC). Methods A retrospective analysis was conducted using the MRI and clinicopathological data of 369 HCC patients who underwent surgery and had pathologically confirmed MVI at the Affiliated Hospital of Qingdao University from January 2019 to September 2024. According to the negative and positive manifestations of MVI, these patients were divided into MVI- group (n=219) and MVI+ group (n=150) and they were then randomly assigned into the training set (n=258) and the test set (n=111) in a ratio of 7:3. Based on the hepatobiliary phase images, the optimal features were extracted and screened from radiomics features, DTL features, and the fusion features of the two. Nine machine learning models were constructed using 3 algorithms (random forest, multi-layer perceptron, and support vector machine, separately) and trained on radiomics features, DTL features, and the fusion features of the two. The diagnostic efficacy of each model was evaluated using receiver operating characteristic curve, and the optimal model was identified as the output model. Results Among all the constructed models, those based on fused features outperformed models using individual features. The random forest classifier model in the training set had the best performance, with an AUC of 0.998 (95% CI: 0.996-1.000), and was therefore selected as the output model in this study. Conclusion Radiomics and DTL models based on gadoxetic acid disodium-enhanced MRI can effectively predict the MVI in HCC. Among these, the random forest classifier model utilizing fused features in the training set exhibits the best performance.
Last Update: 2025-10-29