MRI fusion model classifies endometrial cancer risk in 246-patient study

A retrospective study of 246 patients with pathologically confirmed endometrial carcinoma found that a model combining multiparametric MRI radiomics, 2.5D deep learning features and clinical variables achieved an area under the receiver operating characteristic curve of 0.891 in a 50-patient validation set. The full model reached an accuracy of 0.860, sensitivity of 0.909, specificity of 0.765 and F1 score of 0.896. Researchers used T2-weighted MRI and contrast-enhanced T1-weighted MRI acquired at two Chinese hospitals to estimate whether patients met the 2022 European Society for Medical Oncology definition of high-risk disease before surgery.

Because the cohort was small, retrospective and confined to two institutions, the result supports further investigation rather than clinical use without external validation. Patients who underwent surgery between January 2021 and May 2025 were eligible. Of the 246 participants, 152 came from the Affiliated Hospital of Youjiang Medical University for Nationalities and 94 from Baise City People’s Hospital.

Pathological review classified 90 patients as low risk and 156 as high risk, and the researchers randomly divided them into a training set of 196 and a validation set of 50. The high-risk category included high-grade endometrioid carcinoma, non-endometrioid carcinoma, positive lymphovascular space invasion, positive cervical stromal invasion, positive lymph-node metastasis, or International Federation of Gynecology and Obstetrics stage IB or higher. Patients were excluded if MRI sequences were missing or poor quality, clinical or pathological information was incomplete, they had received preoperative therapy, or another malignancy had been diagnosed concurrently.

Preoperative information included age, body-mass index, carbohydrate antigen 125, 153 and 199 levels, menopausal status, maximum tumor diameter, MRI signs of deep myometrial invasion and cervical stromal invasion, histological type, Ki-67 index, estrogen-receptor status and progesterone-receptor status. Two junior radiologists independently reviewed the images while blinded to the final pathological risk category, and disagreements were resolved with a senior radiologist. MRI had been acquired on three scanner configurations at either 1.5 or 3.0 tesla.

The investigators applied N4 bias-field correction, normalized and resampled images to a uniform 1 × 1 × 1 mm voxel size, and used a fixed gray-level bin width of 100. A radiologist with more than eight years of pelvic MRI experience manually outlined each tumor, and a second senior radiologist reviewed the contours. For radiomic analysis, the team extracted 1,197 quantitative features from each MRI sequence using PyRadiomics.

These comprised shape measurements, first-order intensity statistics, neighborhood gray-tone differences and several texture matrices derived from original images, eight wavelet-decomposed representations and Laplacian-of-Gaussian-filtered images. In parallel, the researchers created 2.5D inputs by cropping the axial slice with the largest tumor cross-section and its adjacent upper and lower slices, resizing each to 224 × 224 pixels and stacking the three images as channels. A ResNet-18 network pretrained on ImageNet supplied 512 deep features from each sequence.

The deep learning analysis therefore used three representative slices rather than the entire tumor volume. Feature selection proceeded in stages. Candidate variables in the training set were standardized, screened with independent-samples t-tests, filtered for correlations above 0.9 and compressed with the least absolute shrinkage and selection operator using 10-fold cross-validation.

The final T2-weighted radiomic model retained 18 features, the contrast-enhanced T1 model retained 13 and the combined-sequence radiomic model retained 10. The corresponding deep learning models retained six, six and 11 features. Random forests were used to build the image-based models, with grid-search hyperparameter optimization and five-fold cross-validation.

A separate clinical logistic-regression model used maximum tumor diameter, MRI evidence of deep myometrial invasion, body-mass index and estrogen-receptor status. The final random forest combined selected radiomic, deep learning and clinical variables. Multiparametric fusion performed better than the single-sequence approaches in the validation analysis.

The T2-weighted radiomic model produced an AUC of 0.701, the contrast-enhanced T1 radiomic model 0.733 and their radiomic fusion 0.764. The corresponding deep learning AUCs were 0.714, 0.740 and 0.771, while the clinical model reached 0.642. The full model’s validation AUC of 0.891 had a 95% confidence interval of 0.800–0.982.

A DeLong test found that, in the training set, the full model performed significantly better than every comparator at P < 0.05. In validation, its advantage was statistically significant over the clinical model and the combined-sequence radiomic model at P < 0.05, but not over the combined-sequence deep learning model, for which P was 0.114. Decision-curve analysis favored the full model’s clinical net benefit, and its calibration curve was described as relatively close to the ideal diagonal.

The researchers reported that the network often concentrated on heterogeneous signal areas within tumors and on lesion margins. These maps provide a view of the model's learned attention but do not establish why a prediction was correct or demonstrate that the heat maps improve clinical decisions. The contrast-enhanced sequence generally outperformed T2-weighted imaging in the single-sequence models.

Together, the results indicate that the engineered radiomic measurements and learned image representations were not identical in their predictive behavior, although the study did not quantify how much unique information each component contributed to the final model. Several findings require caution. The reported association between body-mass index and high-risk disease was opposite to the direction often associated with obesity and endometrial cancer: the multivariate odds ratio was 0.926 per unit of BMI, with P = 0.043.

The authors suggested that low BMI might mark cachexia or poorer systemic status among some patients with advanced tumors, but they described this interpretation as requiring further validation. Histological type could not be estimated reliably in the multivariate model because of complete separation, and progesterone-receptor status was excluded because of strong collinearity with estrogen-receptor status. The 50-patient validation set also makes performance estimates less precise, while the random 8-to-2 split could allow differences between the sets to influence apparent performance.

Generalizability remains the central limitation because the study was retrospective, included only two institutions and did not perform external validation. Images came from scanners with different vendors and magnetic-field strengths, and N4 bias-field correction cannot remove all scanner-dependent signal differences, potentially affecting radiomic reproducibility. Formal inter-reader kappa statistics were not reported for MRI invasion signs despite blinded independent review and consensus.

Only T2-weighted and contrast-enhanced T1-weighted sequences were analyzed, without diffusion-weighted imaging or apparent diffusion coefficient maps. Prospective, multicenter testing would be needed to determine whether the observed discrimination persists across hospitals, scanners, patient populations and clinical workflows. The result is narrower than a demonstration of improved patient care.

Current endometrial-cancer risk stratification relies heavily on postoperative histopathology, including tumor type, invasion depth, lymphovascular involvement and lymph-node status. The study did not test whether its model changed surgical decisions, reduced complications, improved survival or enhanced patient safety. Its supported finding is that, in this retrospective cohort, a random forest combining engineered and learned features from two MRI sequences with clinical variables classified low-risk versus high-risk endometrial carcinoma more accurately than the clinical model alone and most single-approach imaging models.

Whether that advantage survives larger prospective validation and can support preoperative risk assessment remains to be established.

Subject of Research: Preoperative MRI radiomics and deep learning for endometrial carcinoma risk stratification

Article Title: Multiparametric MRI-based radiomics combined with deep learning for preoperative risk stratification of endometrial carcinoma

Article References: Ouyang, Y., Lin, M., Cao, Y., He, N., Zhong, X., & Wu, Y. (2026). Multiparametric MRI-based radiomics combined with deep learning for preoperative risk stratification of endometrial carcinoma. Frontiers in Oncology, 16, Article 1942153. https://doi.org/10.3389/fonc.2026.1942153

Image Credits: AI Generated

DOI: 10.3389/fonc.2026.1942153

Keywords: endometrial carcinoma, MRI, radiomics, deep learning, risk stratification, gynecologic oncology, medical imaging, clinical variables, decision support, Multiparametric, MRI-based, combined