Prediction of Endometrial Cancer Microsatellite Expression Status via Deep Transfer Learning Features Based on Multi-Parameter MRI
10.3969/j.issn.1005-5185.2025.05.017
- VernacularTitle:基于多参数磁共振成像的深度迁移学习特征预测子宫内膜癌微卫星表达状态
- Author:
Ziyan LIU
1
;
Yuxin DING
1
;
Genji BAI
1
Author Information
1. 南京医科大学附属淮安一院影像科,江苏 淮安 223300
- Publication Type:Journal Article
- Keywords:
Endometrial neoplasms;
Magnetic resonance imaging;
Deep transfer learning;
Microsatellite instability;
Forecasting
- From:
Chinese Journal of Medical Imaging
2025;33(5):546-552,561
- CountryChina
- Language:Chinese
-
Abstract:
Purpose To explore deep transfer learning(DTL)based on multi-parameter MRI for joint prediction of uterine endometrial cancer microsatellite expression with clinical parameters.Materials and Methods Retrospective analysis was conducted on 262 patients with uterine endometrial cancer in the Affiliated Huaian NO.1 People's Hospital of Nanjing Medical University from January 2020 to December 2022,who were randomly divided into a training cohort(n=183)and a validation cohort(n=79)in a 7∶3 ratio.DTL features were extracted from T2WI and diffusion-weighted imaging(DWI)using a 50-layer residual neural network(ResNet50).Subsequently,T2-net,DWI-net,multi-sequence fusion and a combined model were constructed.A combined model was constructed via incorporating multiparametric fusion DTL features and clinically independent prognostic factors identified through both univariate and multivariate Logistic regression analysis.Model performance was assessed using the area under the curve,calibration curve and decision curve.Grad-CAM was used for model visualization analysis.Results Compared with single-sequence models,the multi-sequence fusion models exhibited superior performance,with area under the curve value of 0.898 for the validation cohort,accuracy of 0.823,sensitivity of 0.812,specificity of 0.825,respectively,and F1 score was 0.650.Univariate and multivariate Logistic regression analyses revealed that serum human epididymis protein 4 levels and the presence of uterine fibroids were clinically independent risk factors.Ultimately,the combined model demonstrated the best predictive performance in the validation cohort,with area under the curve value of 0.924,accuracy of 0.835,sensitivity of 0.875,specificity of 0.825,respectively,and F1 score was 0.683.Conclusion The multi-parameter fusion model based on DTL features can effectively and non-invasively predict the microsatellite status of endometrial cancer patients.
