Endometrioid adenocarcinoma (EAC) is the primary subtype of endometrial cancer, accounting for 75%-80% of all cases. While often diagnosed at an early stage with favorable outcomes, prognosis varies widely based on several pathological risk factors.
Lymphovascular space invasion (LVSI) - the presence of tumor cells within the lymphatic vessels or small blood vessels just outside the main tumor mass - is one of the most clinically important prognostic factors in EAC. It represents an early step in the process by which cancer cells spread to lymph nodes and distant sites.
LVSI significantly impacts recurrence risk and directly influences recommendations for adjuvant therapy. According to the European Society of Gynecologic Oncology (ESGO) guidelines, LVSI-positive status alone is sufficient to warrant lymph node excision, even without other high-risk histological features.
The critical limitation is that LVSI can only be determined by pathological examination after surgical removal of the uterus. There is currently no validated way to assess LVSI status preoperatively, making surgical planning for lymph node dissection difficult and imprecise.
Radiomics is the extraction of high-dimensional quantitative features from medical images - including texture, intensity, and shape characteristics that are imperceptible to the human eye. These features can encode information about tissue biology and tumor heterogeneity.
While MRI-based radiomics has been extensively studied for predicting LVSI in EC (with AUCs up to 0.89), MRI has practical limitations: it is expensive, time-consuming, and contraindicated in some patients (those with certain implants or claustrophobia). An ultrasound-based approach would be faster, cheaper, and more widely accessible.
Transvaginal ultrasound is already the standard first-line imaging examination for EC, routinely performed in virtually all patients before surgery. Using this existing imaging to predict LVSI status would add significant clinical value without requiring additional diagnostic procedures.
Published studies have demonstrated that ultrasound radiomics can predict LVSI in breast and rectal cancers with AUCs above 0.84. However, no validated ultrasound radiomics model had been developed specifically for LVSI prediction in EAC, representing a gap this study aimed to fill.
This retrospective single-centre study enrolled 171 patients with pathologically confirmed endometrioid adenocarcinoma from December 2017 to December 2022 at the Central Hospital of Wuhan, China. All patients underwent standardized transvaginal ultrasound within 14 days before surgery.
Ultrasound examinations used high-end scanners (Philips iU22, GE VolusonE10, Mindray Resona7) with standardized gynecological presets. Images were captured by a radiologist with over 10 years of gynecological ultrasound experience, selecting the largest cross-section with the clearest visualization of the tumor.
Using the 3D Slicer software platform, two radiologists independently drew regions of interest (ROIs) around the tumor boundary on the ultrasound images. The agreement between radiologists was assessed using the intraclass correlation coefficient (ICC), retaining only features with ICC above 0.75 for downstream analysis.
The study was divided into a training set (119 patients) and test set (52 patients) using stratified random sampling at a 7:3 ratio, ensuring the LVSI-positive to LVSI-negative ratio was maintained in both groups. This preserves the natural class distribution for fair model evaluation.
From each ultrasound image, 837 radiomics features were extracted using the 3D Slicer Radiomics Extension Pack, encompassing first-order statistics (intensity distributions), shape descriptors, and texture features including wavelet-transformed image metrics.
Feature selection used a three-step reduction pipeline: Z-score normalization, statistical t-tests retaining features with p < 0.05 (yielding 152 features), Pearson correlation filtering to remove redundant features with correlation above 0.9 (yielding 48 features), and finally LASSO regression with 10-fold cross-validation to select the final seven predictive features.
The seven selected features were primarily from wavelet-transformed images and included both first-order statistics (capturing intensity distribution) and texture features (capturing spatial heterogeneity patterns). Wavelet transformation decomposes images into different frequency components, revealing patterns at multiple scales simultaneously.
A Rad-score (radiomics score) was constructed as a weighted linear combination of the seven selected features, providing a single numerical value for each patient that summarizes their imaging-based LVSI risk signature.
Among 171 patients, 39 (22.8%) were LVSI-positive. Univariate analysis identified preoperative histological grade and Ki-67 as significant predictors of LVSI. On multivariate analysis, only histological grade remained an independent clinical predictor.
The ultrasound radiomics model alone achieved an AUC of 0.80 (95% CI: 0.71-0.89) in the training set and 0.74 (95% CI: 0.56-0.92) in the test set - demonstrating that imaging features can meaningfully predict LVSI beyond what is visible to the eye during standard ultrasound interpretation.
Combining the Rad-score with histological grade in a comprehensive prediction model improved performance: AUC 0.84 (95% CI: 0.76-0.92) in training and 0.75 (95% CI: 0.57-0.93) in the test set. A nomogram was constructed to visualize how these two variables combine to estimate individual LVSI probability.
Critically, the combined model achieved a sensitivity of 96% in the training set and 75% in the test set, with specificities of 64% and 70% respectively. The high sensitivity is clinically meaningful for a screening purpose - it means the model misses very few LVSI-positive patients who might benefit from lymph node assessment.
A decision curve analysis (DCA) confirmed that the combined nomogram has strong clinical utility across a range of threshold probabilities. DCA evaluates whether using a model to make treatment decisions produces a net clinical benefit compared to treating all patients or no patients.
The calibration curve demonstrated good agreement between predicted and observed LVSI probabilities, confirming that the model's numerical probability estimates are accurately calibrated - not just directionally correct but quantitatively reliable.
The high sensitivity of the model (prioritizing detection over specificity) is appropriate for its intended clinical use: identifying LVSI-positive patients who need comprehensive lymph node staging. Missing an LVSI-positive patient (false negative) has more serious consequences than a false positive recommendation for staging.
Preoperative knowledge of LVSI status could help surgeons plan the extent of lymph node dissection more precisely - avoiding unnecessary extensive dissection in low-risk patients while ensuring high-risk patients receive complete staging, reducing both over-treatment and under-treatment.
This study establishes the first validated ultrasound-based radiomics model for LVSI prediction in endometrioid adenocarcinoma, filling an important gap in preoperative risk assessment tools for endometrial cancer.
The clinical appeal is significant: ultrasound is already performed in virtually all EC patients before surgery, requires no additional testing or imaging costs, and is available at most gynecological oncology centers globally - including in resource-limited settings where MRI may not be accessible.
Combining the objective, quantitative Rad-score with the established clinical predictor of histological grade produces a model that integrates pathological tissue information with imaging information, reflecting the multidimensional nature of LVSI risk.
Future work should validate this model in larger multicenter cohorts to confirm generalizability, explore whether other clinical variables or dynamic ultrasound features (such as elastography or contrast-enhanced ultrasound) can further improve predictive accuracy beyond the current performance.