Endometrial cancers are broadly divided into two groups. Type I cancers are estrogen-dependent, typically low-grade, and carry a favorable prognosis. They account for 80-90% of all cases and are associated with endometrial hyperplasia. Type II cancers are estrogen-independent, tend to be high-grade and aggressive, and make up the remaining 10-20% of cases.
Type II endometrial cancer includes several subtypes: grade 3 endometrioid adenocarcinoma, serous adenocarcinoma, clear cell carcinoma, and carcinosarcoma. These tumors frequently spread to distant organs by the time of diagnosis and are responsible for over 50% of all endometrial cancer deaths, despite being the minority of cases.
Distinguishing between Type I and Type II before surgery is critically important because surgical and adjuvant treatment strategies differ significantly between the two. Type II patients typically require more aggressive surgery and additional therapies like chemotherapy and radiation, while Type I patients may be candidates for more conservative approaches.
Currently, the standard preoperative diagnostic method is dilatation and curettage (D&C), an invasive procedure that samples tissue from the uterine cavity. However, D&C has limitations: sampling errors can occur, and it cannot assess depth of invasion or lymph node involvement, which are key factors in treatment planning.
Radiomics is the process of extracting large numbers of quantitative mathematical features from medical images. Rather than a radiologist making a subjective judgment about a scan, radiomics algorithms calculate hundreds of measurable properties like texture patterns, signal intensities, and shape characteristics from the tumor region.
For this study, features were extracted from three MRI sequences: T2-weighted imaging (T2WI), which shows tissue anatomy; apparent diffusion coefficient (ADC) maps, which measure how freely water molecules move through tissue (restricted in densely packed cancer cells); and dynamic contrast-enhanced MRI at the delayed phase (DCE4), which shows how blood flow changes after a contrast dye is injected.
Combining multiple imaging sequences - called multiparametric MRI - provides more information about tumor biology than any single scan alone. Each sequence highlights different aspects of tumor behavior: cell density, blood supply, and tissue architecture.
A total of 720 radiomics features were extracted from each MRI sequence for each patient. After statistical selection using LASSO regression to identify the most informative and non-redundant features, 12 key features from combined ADC and DCE4 sequences were ultimately used to build the final predictive model.
The study enrolled 403 patients with histopathologically confirmed endometrial cancer from two hospitals in Nanjing, China. Of these, 316 had Type I and 87 had Type II cancer. Patients were randomly split 70/30 into training and validation cohorts, following a design meant to evaluate whether models generalize to new patients.
To find the best predictive approach, six different machine learning algorithms were tested: logistic regression, random forest, bootstrap aggregating (Bagging), support vector machine (SVM), artificial neural network (ANN), and naive Bayes (NB). Each algorithm was applied to seven different combinations of MRI feature sets, resulting in 42 distinct models.
In addition to imaging features, the model also incorporated biomarkers from preoperative biopsy specimens: Ki-67 (a measure of tumor cell division rate), p53 protein expression (indicating mutation status), and estrogen receptor (ER) expression. These molecular markers were obtained from the less invasive curettage procedure rather than requiring surgery.
A nomogram was then constructed by combining the best radiomics model with the significant clinical and molecular features. A nomogram is a graphical tool that allows clinicians to estimate an individual patient's probability of having Type II cancer by reading off scores from multiple scales corresponding to different predictors.
Among all 42 radiomics models tested, the Naive Bayes (NB) classifier using 12 features from ADC and DCE4 sequences performed best. It achieved an AUC (Area Under the Curve) of 0.927 in training and 0.869 in validation. AUC ranges from 0.5 (no better than chance) to 1.0 (perfect prediction), so these values indicate strong discriminative ability.
Key molecular markers that independently predicted Type II cancer included p53 mutation status, elevated Ki-67 expression, and lower estrogen receptor (ER) expression. Among conventional MRI features, lower ADC values, greater tumor enhancement on DCE4, and the presence of lymph node involvement were significant predictors.
The integrated nomogram combining radiomics, imaging, and molecular markers outperformed any single predictor in both training (AUC = 0.951) and validation (AUC = 0.915) sets. For comparison, p53 alone achieved only AUC = 0.683 in validation, highlighting the value of combining multiple data sources.
Decision curve analysis (DCA) showed that the nomogram provided a high net clinical benefit across a wide range of risk thresholds, meaning the model would help more patients receive appropriate care than decisions based on any single test alone.
ADC (apparent diffusion coefficient) measures how freely water moves through tissue. Type II endometrial cancers tend to have lower ADC values than Type I cancers, reflecting their higher cellular density. More cells packed together physically restrict water movement, producing a lower diffusion signal.
DCE4 (delayed phase contrast enhancement) measures how blood flow changes over time after contrast injection. Type II tumors showed significantly higher enhancement rates, suggesting more abnormal, rapidly growing tumor blood vessels. This difference in vascular behavior reflects the more aggressive biology of high-grade cancers.
The study's multicenter design and use of multiple MRI sequences are strengths that overcome limitations of prior single-center, single-sequence studies. However, the relatively small number of Type II cases (87 out of 403) is a limitation that could affect model reliability and generalizability to populations with different Type I/II proportions.
The authors note that future work should incorporate additional IHC markers such as p16 and WT-1, explore automated tumor segmentation to reduce radiologist variability, and compare radiomics models against deep learning approaches on larger multicenter datasets with more balanced histological type distributions.
The nomogram could change how patients are worked up before surgery. Instead of relying solely on D&C biopsy results, which can miss the true tumor grade, clinicians could use the combined MRI radiomics and molecular marker nomogram to estimate the probability that a patient has aggressive Type II disease.
Patients predicted to have Type II cancer could be directed toward more comprehensive staging surgery and earlier discussion of adjuvant therapies, without waiting for final pathology after the operation. This could prevent under-treatment of high-risk patients who might otherwise be managed like typical low-grade cases.
Conversely, patients with low nomogram scores could potentially avoid unnecessarily aggressive surgery. Over-treatment in low-risk patients carries real harms including surgical complications, lymphedema from lymph node removal, and reduced quality of life.
Because the model uses preoperative MRI that patients already undergo as part of standard staging workup, implementation would not require additional invasive procedures. The radiomics analysis could be added as a software layer on top of existing imaging infrastructure, making clinical adoption potentially straightforward.