Related Experiment Video
Updated: Jun 24, 2026
Murine Model for Non-invasive Imaging to Detect and Monitor Ovarian Cancer Recurrence
Published on: November 2, 2014
Development and Validation of a Multi-Omics Model Integrating US-Derived and WSI-Based Features to Predict Lymph Node
Ge Yan1, Xiujuan Wu2, Peiting Zhao3
1State Key Laboratory of Radiation Medicine and Protection, School of Radiation Medicine and Protection, Collaborative Innovation Center of Radiological Medicine of Jiangsu Higher Education Institutions, Soochow University, Suzhou, 215000, People's Republic of China.
Objective:
The study aims to develop and validate a multi-omics model based on preoperative ultrasound (US) imaging results, intraoperative H&E- stained slides, and clinical features to predict lymph node metastasis (LNM) before lymph node dissection (LND) in ovarian cancer (OC) patients.
Methods:
We analyzed 157 OC patients undergoing LND with definitive pathological confirmation of LNM status, comprising 91 patients in the training cohort, 38 in the internal validation cohort, and 28 in the external test cohort. US images were processed with PyRadiomics to extract radiomics features, while pathological WSIs were processed with deep learning (DL) algorithms and multi-instance learning(MIL) algorithms to extract pathomics features. Then, radiomics and pathomics models were developed using support vector machines (SVMs), logistic regression (LR), and extreme gradient boosting (XGBoost) after dimensionality reduction and feature selection. To create a powerful multi-omics model, clinical features were incorporated into the optimal radiomics and pathomics features. Performance of models was assessed by accuracy, AUC, 95% CI, sensitivity, specificity, PPV and NPV.
Results:
A total of 11 features were used to build radiomics models out of a selection of 1561 radiomics features. The SVM_rad model demonstrated superior predictive performance (AUC: training=0.816, validation=0.760, test=0.775). In parallel, pathomics models were built using a refined set of 3 features selected from the original 206 pathomics features. Among these, the SVM_path model showed the highest predictive efficiency (AUC: training=0.983, validation=0.817, test=0.813). The multi-omics model showed the greatest discriminative power (AUC: training=0.988; validation=0.923; test cohort=0.862). The quality of the prediction model was demonstrated by the DeLong test, calibration curves, and decision curve analysis, which verified its high discrimination, calibration, and clinical usefulness.
Conclusion:
The study's findings indicate that the multi-omics model integrating the tumor-level radiological data, cellular-level pathological information, and patient-level clinical features can predict LNM before LND in OC and support rational treatment plans.
