Predicting microvascular invasion in hepatocellular carcinoma with multi-sequence MRI and a Swin Transformer-based deep learning model(PDF)
《中国医学物理学杂志》[ISSN:1005-202X/CN:44-1351/R]
- Issue:
- 2026年第2期
- Page:
- 245-254
- Research Field:
- 医学人工智能
- Publishing date:
Info
- Title:
- Predicting microvascular invasion in hepatocellular carcinoma with multi-sequence MRI and a Swin Transformer-based deep learning model
- Author(s):
- HUANG Qian1; ZHUANG Yinping1; XU Peng2; GONG Ping1
- 1. School of Medical Imaging, Xuzhou Medical University, Xuzhou 221004, China 2. Department of Radiology, the Affiliated Hospital of Xuzhou Medical University, Xuzhou 221000, China
- Keywords:
- Keywords: hepatocellular carcinoma microvascular invasion multi-sequence magnetic resonance imaging Swin Transformer radiomics
- PACS:
- R318；R735.7
- DOI:
- DOI:10.3969/j.issn.1005-202X.2026.02.015
- Abstract:
- Abstract: Objective To develop and validate a Swin Transformer (ST)-based deep learning (DL) model using multi-sequence magnetic resonance imaging (MRI) for predicting microvascular invasion (MVI) in hepatocellular carcinoma (HCC), thereby providing a novel solution and objective scientific evidence to facilitate the precision diagnosis and treatment for HCC patients. Methods A retrospective analysis was performed on 174 patients with surgically and pathologically confirmed HCC who were admitted to the Affiliated Hospital of Xuzhou Medical University. Preoperative multi-sequence MRI images, including arterial phase, delayed phase (DP), fat-suppressed T2-weighted imaging, and diffusion-weighted imaging, were collected for each patient. After image preprocessing and enhancement, an ST-based DL model was established and then compared with models based on convolutional neural network architectures, including DenseNet121, DenseNet169, ResNet34, ResNet50, VGG16, and GoogleNet, as well as a Transformer-based Vision Transformer model and a radiomics model. Model accuracy, area under the receiver operating characteristic curve (AUC), sensitivity, and specificity were calculated, and the model performance was evaluated using confusion matrices and decision curve analysis. To enhance model interpretability, gradient-weighted class activation mapping (Grad-CAM) was employed to visualize the regions of interest focused on by the models. Results The ST model demonstrated the optimal overall performance among the 4 sequences in MVI prediction, with particularly outstanding results on the DP sequence. Specifically, it achieved an accuracy of 0.944, an AUC of 0.993 (95% CI: 0.985-0.998), a sensitivity of 0.984, and a specificity of 0.904, significantly outperforming other DL models and the radiomics model. Decision curve analysis further confirmed its superior potential for clinical application. Grad-CAM visualization indicated that the ST model effectively concentrated on tumor-associated regions. Conclusion The ST model exhibits excellent predictive performance across multiple MRI sequences and can served as a powerful auxiliary tool to support clinical diagnosis, treatment decision-making, and prognostic evaluation.
Last Update: 2026-01-27