Endometrial cancer treatment decisions depend heavily on risk classification. Low-risk patients (early stage, low grade, superficial invasion) can be treated with straightforward hysterectomy, while high-risk patients require more extensive surgery including lymph node dissection, and often adjuvant chemotherapy or radiation. Incorrectly classifying risk leads to either over-treatment (unnecessary surgery and its complications) or under-treatment (inadequate cancer control).
Current preoperative classification relies on clinical factors - age, tumor size, imaging findings - and biopsy results showing cancer grade and histological type. However, final pathological staging often reveals features not apparent preoperatively, such as deep myometrial invasion or lymph node involvement. This disconnect between preoperative prediction and actual surgical findings drives interest in better imaging-based risk tools.
This study focused specifically on endometrial endometrioid adenocarcinoma (the most common subtype, arising from the glandular cells of the uterine lining) and developed a model combining traditional radiomics, deep learning features, and clinical data extracted from MRI to classify patients as low or high surgical risk before their operation.
The study included 168 patients with histologically confirmed endometrial endometrioid adenocarcinoma, split into a training cohort of 95 patients and a validation cohort of 73 patients. All patients underwent preoperative MRI including standard sequences and diffusion-weighted imaging (DWI), which measures how freely water molecules move in tissue and provides information about tumor cellularity.
Traditional radiomics involved manual segmentation of the tumor on MRI followed by automated extraction of hundreds of quantitative features - including shape metrics, texture measures, and signal intensity statistics. Machine learning (LASSO regression) was then used to select the most predictive features and build a radiomics score.
Deep learning features were extracted using a convolutional neural network applied directly to the MRI images, learning complex visual patterns without requiring manual feature definition. Finally, clinical features including age, BMI, CA125 (a blood cancer biomarker), and tumor size were included. A nomogram - a visual prediction tool combining multiple model inputs - was constructed to generate individual patient risk scores.
The combined radiomics nomogram achieved strong performance in the training cohort, with an AUC of 0.923. More importantly, this performance held up well in the independent validation cohort, where the AUC was 0.842. While there is some drop from training to validation (which is expected), an AUC above 0.84 in an independent cohort is considered clinically meaningful performance.
The combined model outperformed each individual component. Traditional radiomics alone achieved AUC 0.86 in training; deep learning features alone achieved 0.81; clinical variables alone achieved 0.79. The combined model's superior performance confirms the value of integrating complementary information sources - each captures different aspects of tumor biology that the others miss.
Calibration analysis (using Hosmer-Lemeshow test and calibration curves) confirmed that the model's predicted probabilities matched actual risk levels well, meaning the model doesn't just rank-order patients but also gives accurate probability estimates. Decision curve analysis showed net clinical benefit across a range of risk thresholds, demonstrating that using the model for clinical decisions would genuinely help more patients than harm them compared to treating everyone or no one.
Feature importance analysis revealed that diffusion-weighted imaging (DWI)-derived features contributed substantially to model performance, particularly the apparent diffusion coefficient (ADC) - a measure of how restricted water movement is within the tumor. High-grade, invasive tumors typically show lower ADC values due to their dense cellular packing, and this study confirmed ADC as one of the strongest individual predictors of high-risk pathology.
Among traditional radiomic features, texture features capturing tumor heterogeneity were most informative. Tumors with irregular, non-uniform internal texture patterns were more likely to have deep myometrial invasion and lymphovascular involvement. This reflects the biological reality that high-risk tumors often have heterogeneous histological composition, with areas of varying differentiation, necrosis, and invasion.
Clinical variables that contributed most were tumor size and CA125. Larger tumors and elevated CA125 are known prognostic factors, and their inclusion in the model reflects the principle that well-established clinical knowledge should be integrated with novel imaging features rather than replaced by them.
The practical application of this nomogram is straightforward: before surgery, a patient's MRI is acquired, the tumor is segmented, features are extracted, clinical values are entered, and the nomogram outputs a risk score with a predicted probability of being high-risk. Surgeons can use this probability to decide whether to plan a more extensive procedure with lymph node dissection from the outset.
This matters particularly for patients who might appear low-risk based on clinical and biopsy information alone but who have imaging features suggesting occult deep invasion. Identifying these patients preoperatively could prevent the need for re-operation - a second surgery when final pathology reveals unexpected high-risk features not acted upon during the initial operation.
The nomogram format also facilitates shared decision-making: the probability score can be shown to patients in a way that contextualizes their individual risk, rather than just placing them in a binary low/high-risk category. This supports informed consent and helps patients understand why a more extensive or less extensive operation is being recommended.
This study contributes to a growing evidence base that combined radiomic-deep learning models outperform traditional imaging interpretation for endometrial cancer risk classification. The approach is particularly valuable because it extracts quantitative, objective information from MRI examinations that are already routinely obtained, adding new value without additional patient burden or cost.
Key limitations include the single-center design and relatively modest sample size. The model was developed and validated in Chinese patients at one institution, and the specific radiomic features selected may reflect local MRI protocol parameters. Validation in external cohorts with different MRI equipment and protocols is essential before widespread adoption.
The study also highlights a methodological advance in the field: the explicit combination of three complementary feature types (traditional radiomics, deep learning, clinical) into a single unified model. This framework - rather than the specific features identified - may be the most generalizable contribution, providing a template for similar combined-modality prediction models in other cancer types where preoperative risk stratification is clinically important.