Although many endometrial cancer patients do well after surgery, a meaningful fraction experience disease recurrence and shortened survival. Predicting which patients are at highest risk of dying from their cancer - and understanding why - could allow doctors to tailor treatment more aggressively for high-risk individuals while sparing low-risk patients from unnecessary side effects.
Current prognostic tools rely mainly on surgical pathology findings like tumor stage, grade, and type. These are powerful predictors but are only available after surgery. A radiomics-based model built from preoperative MRI scans could provide prognostic information before any procedure, enabling better initial treatment planning.
The study enrolled 602 endometrial cancer patients from three separate medical centers in China. Using this multicenter design helps ensure that the model works across different patient groups and imaging equipment rather than just in one setting. Patients were followed up to track whether they survived over 1, 3, and 5 years after diagnosis.
Radiomics features were extracted from two regions in the MRI scans: the tumor itself and the area immediately surrounding the tumor (called the peritumoral region). The peritumoral region reflects changes in normal tissue caused by the cancer growing nearby. An XGBoost machine learning model was trained combining features from both regions. LASSO regression was used to select the most informative features and prevent overfitting.
The combined tumor plus peritumoral radiomics model performed well across all testing datasets. In the validation cohort, the model achieved AUCs of 0.862, 0.885, and 0.870 for predicting 1-year, 3-year, and 5-year overall survival respectively. On two independent external test sets, AUCs ranged from 0.731 to 0.869, demonstrating that the model retained predictive power on entirely new patients.
Importantly, the combined model consistently outperformed models built from tumor features alone, confirming that information from the peritumoral region carries additional prognostic value. This makes biological sense - the tissue microenvironment surrounding a tumor influences how aggressively it behaves and how likely it is to spread.
Beyond predicting survival, the researchers explored why radiomics features correlate with outcomes by connecting imaging patterns to gene expression data. They identified FLT1 - a gene involved in blood vessel formation (angiogenesis) - as a key biological link. The correlation between the radiomics score and FLT1 expression was very strong (r=0.85, P=0.001).
Tumors with high FLT1 activity grow new blood vessels more aggressively, which fuels tumor growth and spread. When FLT1 expression is high, patients tend to have worse survival. This finding suggests that the radiomics model is not just capturing arbitrary image patterns but is detecting real biological signals related to tumor aggressiveness and blood supply.
The researchers went further by combining the radiomics model with additional biological data layers - including gene expression (transcriptomics) and protein data (proteomics) - in what is called a multi-omics analysis. When all data layers were combined, the model achieved near-perfect survival prediction: AUCs of 0.989, 0.996, and 1.000 for 1-year, 3-year, and 5-year survival.
While such extremely high accuracy in a research setting does not mean perfect real-world performance, these results highlight a key principle: combining imaging information with molecular biology data dramatically improves our ability to understand and predict individual patient outcomes. This kind of integrated approach points toward the future of precision oncology, where treatment decisions are guided by a comprehensive picture of each patient's tumor.
This study demonstrates that a preoperative radiomics model built from MRI scans can meaningfully predict endometrial cancer survival and connects those predictions to real biological mechanisms. The multicenter design strengthens confidence that the model has genuine generalizability rather than just working in one specialized hospital setting.
For the model to reach clinical use, further prospective validation is needed - testing it on new patients in real-time rather than looking back at historical data. Integration with other standard clinical variables and compatibility with different MRI systems also needs to be confirmed. Still, this work represents a significant step toward personalized risk stratification in endometrial cancer that begins before surgery, when treatment decisions matter most.