MRI-based deep learning models for quantitative prediction of occult peritoneal metastasis in advanced gastric cancer(PDF)
《中国医学物理学杂志》[ISSN:1005-202X/CN:44-1351/R]
- Issue:
- 2026年第7期
- Page:
- 980-986
- Research Field:
- 医学人工智能
- Publishing date:
Info
- Title:
- MRI-based deep learning models for quantitative prediction of occult peritoneal metastasis in advanced gastric cancer
- Author(s):
- YAO Chun; XIONG Bo; YANG Guihan; LIANG Yuemei; LI Jianhui; YANG Zhiqi
- Medical Imaging Center, Meizhou Peoples Hospital, Meizhou 514031, China
- Keywords:
- Keywords: gastric cancer occult peritoneal metastasis magnetic resonance imaging deep learning
- PACS:
- R318；R735.2
- DOI:
- DOI:10.3969/j.issn.1005-202X.2026.07.020
- Abstract:
- Abstract: Objective To explore the feasibility and value of multiple deep learning models based on magnetic resonance imaging (MRI) for the prediction of occult peritoneal metastasis (OPM) in advanced gastric cancer. Methods A retrospective analysis was conducted on 476 patients with pathologically confirmed advanced gastric cancer and negative preoperative MRI findings for peritoneal metastasis. All patients were randomly divided into a training set (n=334) and a test set (n=142) at a ratio of 7:3. Three-dimensional tumor regions of interest were manually delineated on T2WI-FS sequences. Three deep learning models, including ResNet50, Med-ViT, and TP-Mamba, were trained, and their performance was evaluated using the area under the receiver operating characteristic curve (AUC), sensitivity, specificity, the DeLong test, integrated discrimination improvement, and decision curve analysis. Results Med-ViT model (training set AUC=0.860 test set AUC=0.763) exhibited superior performance in predicting OPM compared with TP-Mamba model (training set AUC=0.767 test set AUC=0.652) and ResNet50 model (training set AUC=0.528 test set AUC=0.570). In both the training and test sets, the DeLong test showed that Med-ViT had significantly higher AUC values than the other two models (P<0.05). Integrated discrimination improvement analysis demonstrated that Med-ViT model achieved the optimal comprehensive discrimination improvement ability, and decision curve analysis reflected its greatest net benefit. Conclusion The self-attention mechanism-based Med-ViT deep learning model can accurately predict the risk of OPM in advanced gastric cancer and outperforms ResNet50 and TP-Mamba. Therefore, it is expected to serve as a reliable auxiliary tool for non-invasive preoperative screening of patients at high risk of OPM.
Last Update: 2026-07-23