Multi-scale radiomics combined with deep learning for pancreatic cancer prognosis prediction: model construction and validation
10.3760/cma.j.cn113884-20250328-00103
- VernacularTitle:多尺度影像组学融合深度学习胰腺癌预后预测模型的构建和评估
- Author:
Yixuan SHEN
1
;
Chengwei CHEN
1
;
Wenbin LIU
1
;
Xinyue ZHANG
1
;
Yun BIAN
1
;
Chengwei SHAO
1
Author Information
1. 海军军医大学第一附属医院放射诊断科，上海 200433
- Publication Type:Journal Article
- Keywords:
Pancreatic neoplasms;
Prognosis;
Deep learning;
Radiomics
- From:
Chinese Journal of Hepatobiliary Surgery
2025;31(9):678-684
- CountryChina
- Language:Chinese
-
Abstract:
Objective:A prognosis prediction model for pancreatic cancer was constructed based on multi-scale radiomics combined with deep learning, and the prediction effect of the model was evaluated.Methods:A retrospective analysis was conducted on the clinical data of 215 patients who underwent radical resection of pancreatic cancer at the First Affiliated Hospital of Naval Medical University from January 2017 to December 2017. Among them, 134 were male and 81 were female, with an age of (61.9±9.2) years. Patients were randomly divided into the training set ( n=151) and the test set ( n=64) in a ratio of 7: 3. Habitat features, peritumoral radiomics features, 3D radiomics features, and 2.5D deep learning features were extracted from preoperative CT images respectively. After feature screening, a survival prediction model was constructed using the CoxBoost machine learning algorithm that integrated the Boosting algorithm and the Cox proportional hazards model. The performance of the model was evaluated using the area under the time-dependent receiver operating characteristic curve and the consistency index. The clinical benefits of the model were evaluated using decision curve analysis. The survival curves were plotted using the Kaplan-Meier method, and the log-rank test was used for the comparison of survivals between groups. Results:The LASSO, random forest and extreme gradient boosting models were each used to screen out the top 10 most important features and take the union, ultimately obtaining 20 radiomics features for modeling. In the training set and test set, the consistency index of the CoxBoost model in predicting overall survival was 0.717 (95% CI: 0.669-0.765) and 0.688 (95% CI: 0.610-0.766), respectively, and the area under the curve for predicting overall survival at 1, 2, and 3 years after surgery was 0.830 (95% CI: 0.752-0.898), 0.753 (95% CI: 0.665-0.833), 0.828 (95% CI: 0.735-0.908) and 0.690 (95% CI: 0.549-0.824), 0.780 (95% CI: 0.649-0.887 and 0.793 (95% CI: 0.660-0.897), respectively. The area under the curve for predicting long-term survival after surgery (≥40 months) was above 0.8. Based on the optimal cutoff value of -0.19 for the predicted value of the CoxBoost model calculated by the R package " survminer", the patients were divided into high-risk (predicted value >-0.19) and low-risk (predicted value <-0.19) groups. In both the training set and the test set, the survival of patients in the low-risk group was better than that in the high-risk group (training set: χ2=39.01, P<0.001; test set: χ2=12.34, P<0.001). The median survival period of patients in the high-risk group was lower than that in the low-risk group (training set: 15.80 vs 34.07 months; test set: 16.87 vs 43.07; months). Decision curve analysis shows that patients obtain survival benefit when the threshold probability of the training set is greater than 0.25 and that of the test set is greater than 0.45. Conclusion:The CoxBoost model has a good predictive ability for the overall survival of pancreatic cancer patients after surgery and can effectively screen out patient subgroups that may significantly benefit from surgical treatment.
