Rising global incidence and disparity. Endometrial cancer (EC) is the most prevalent gynecological malignancy in high-income countries, with 417,367 new cases reported worldwide in 2020. Annual US diagnoses are projected to double by 2030 to approximately 122,000 cases. Significant racial disparities exist: Black women are more frequently diagnosed with aggressive EC subtypes and advanced stages, consistently leading to poorer outcomes than White women.
EC subtypes and their imaging implications. Type I EC (approximately 80% of cases) is estrogen-dependent, well-differentiated, associated with PTEN mutations in 30%-80% of cases, and carries a favorable prognosis. Type II EC (approximately 20%) arises in older women, is poorly differentiated, frequently p53-mutated, and is associated with early lymphatic spread and aggressive behavior. These biological differences have direct implications for imaging strategy and staging accuracy.
Clinical presentation and early detection imperative. Postmenopausal bleeding (PMB) prompts approximately two-thirds of gynecological consultations in peri- and post-menopausal women and is present in nearly 90% of EC patients, yet results in cancer in only 9% of cases. Five-year survival exceeds 95% when caught early but falls to approximately 15% after distant spread, making accurate triage of PMB patients critically important.
Imaging modalities overview. TVUS is the first-line imaging tool due to its accessibility and cost-effectiveness - an endometrial thickness above 5 mm in postmenopausal women yields approximately 96% sensitivity and 61% specificity. MRI is the gold standard for local staging due to its superior soft tissue contrast. CT is reserved for detecting distant metastases. The integration of AI into ultrasound imaging is now expanding diagnostic capabilities beyond standard 2D TVUS.
Search strategy. A systematic literature search was conducted across PubMed, Scopus, Web of Science, and Google Scholar for studies published between January 2010 and March 2025. Search terms combined "endometrial cancer" with imaging modalities including "transvaginal ultrasound", "3D ultrasound", "contrast-enhanced ultrasound", "elastography", "artificial intelligence", "radiomics", and "deep learning".
Inclusion criteria and quality assessment. Eligible studies evaluated TVUS, 3D-US, CEUS, elastography, or AI-enhanced US for EC diagnosis or staging, involved human subjects, and were published in English. Methodological quality was assessed using the QUADAS-2 tool across four domains: patient selection, index test conduct, reference standard quality, and flow and timing. Out of 41 included studies, most demonstrated low risk of bias, though elastography and AI studies showed more variable rigor.
Performance synthesis. Sensitivity, specificity, and AUC values were extracted from eligible studies. Formal meta-analysis was not feasible due to substantial heterogeneity in study designs, patient populations, and imaging protocols. A comparative summary table was developed displaying pooled or median performance metrics for each modality, along with their primary indications, strengths, and limitations.
Sources of heterogeneity. Substantial variation in reported performance across studies was attributed to differences in patient demographics (menopausal status, BMI), imaging equipment, operator expertise, diagnostic thresholds, and reference standards (histopathological confirmation vs. MRI-only staging). These factors underscore the need for standardized imaging protocols and multicenter external validation.
TVUS diagnostic performance. TVUS exhibited high sensitivity (76%-96%) for initial EC screening but moderate to variable specificity (61%-86%), with a pooled AUC of 0.88. Performance was particularly variable in postmenopausal and obese patients. An endometrial thickness cutoff of 5.15 mm yielded sensitivity 80.5% and specificity 86.2% for detecting EC or atypical endometrial hyperplasia in one prospective study. TVUS is limited in its ability to assess myometrial invasion depth and cervical stromal involvement.
MRI and 3D-US staging accuracy. MRI provided superior specificity (84%-95%) with sensitivity 79%-92% and AUC 0.89-0.91, excelling in assessing myometrial invasion depth, cervical stromal involvement, and type II EC subtypes. 3D-US achieved diagnostic accuracy approaching MRI in carefully selected early-stage patients, with sensitivity 75%-88%, specificity 75%-91%, and AUC 0.86-0.90. Endometrial volume assessment by 3D-US showed higher specificity (75%) than 2D endometrial thickness measurement (69%).
CEUS and elastography. CEUS demonstrated sensitivity 82%-90%, specificity 78%-89%, and AUC 0.85-0.88, providing real-time microvascular imaging that distinguishes malignant tumors (rapid, uneven, hypervascular enhancement) from benign lesions. Shear wave elastography (SWE) showed sensitivity 78%-85%, specificity 70%-88%, and AUC 0.83-0.87; malignant lesions showed significantly higher mean SWE values (17.14 kPa) than benign (11.49 kPa) and normal tissues (10.39 kPa).
AI-enhanced ultrasound. AI-enhanced US models incorporating radiomics or deep learning demonstrated sensitivity 80%-90%, specificity 78%-92%, and pooled AUC up to 0.91, particularly for risk prediction and lesion segmentation tasks. XGBoost applied to clinical and US features for thyroid carcinoma classification achieved AUC 0.969 in one representative study. Deep learning models for US image segmentation achieved Dice similarity coefficients above 0.91.
Why TVUS and MRI are complementary rather than interchangeable. While pooled analyses show TVUS sensitivity of 76%-96% and MRI sensitivity of 79%-92% are comparable for detecting myometrial invasion, this equivalence holds primarily for early-stage, low-risk disease. MRI remains unequivocally superior for cervical stromal invasion assessment, lymph node evaluation, and type II EC detection - all critical for FIGO staging and surgical planning.
Proposed tiered diagnostic pathway. The authors propose a three-step algorithm: TVUS for initial screening due to its affordability and accessibility; MRI for intermediate-to-high-risk cases when TVUS suggests structural distortion, myometrial invasion, or possible type II histology; and CT or PET-CT for distant metastasis evaluation in suspected advanced disease. In resource-limited settings, AI-enhanced US could serve as a surrogate for MRI.
Subtype-specific imaging considerations. Type II EC, including serous, clear cell, and high-grade subtypes, often presents with minimal endometrial thickening and flat morphology that is challenging to detect by US alone. These tumors are frequently p53-mutated and may exhibit HER2 overexpression, requiring MRI for reliable assessment of deep myometrial and cervical stromal invasion, lymphovascular involvement, and adnexal spread.
AI limitations and future requirements. Most AI models evaluated lack robust external validation across diverse clinical environments. The lack of standardized image acquisition protocols complicates AI model training across institutions. Regulatory approval requires evidence of safety, reproducibility, and clinical utility that many academic prototype models still lack. AI models also raise data privacy, training dataset bias, and health equity concerns particularly for deployment in low-income and middle-income countries.
Imaging-guided surgical planning. Accurate assessment of myometrial invasion depth directly determines the need for lymphadenectomy, influencing surgical morbidity and survival outcomes. Radiomics-enhanced US can assist in preoperative risk stratification, helping identify patients who may benefit from extended surgical staging. Sequential diagnostic protocols using TVUS initially and reserving MRI for ambiguous or high-risk cases optimize both resource allocation and diagnostic accuracy.
AI integration into real-time ultrasound. AI integration into real-time US - including automated endometrial volume measurement, tumor boundary segmentation, and predictive risk scoring from radiomic features - is approaching clinical readiness in high-resource centers. Hand-held AI-enhanced US devices hold particular promise for bedside diagnostics and triage in resource-limited settings and rural or underserved healthcare systems.
Health equity considerations. MRI remains inaccessible in many low-resource settings due to high costs and logistical limitations. AI-enhanced portable US platforms offer a feasible and scalable alternative, particularly when embedded with cloud-based decision support. Ensuring diverse, representative training datasets and involving local clinical stakeholders in tool design is essential for closing diagnostic gaps across populations and avoiding widening existing disparities.
Personalized medicine and radiomics. The combination of ML and radiomics extracted from US images provides a quantitative, repeatable approach to tissue characterization beyond human visual perception. Future models could identify high-risk molecular features such as p53 mutation status or non-endometrioid histology through advanced pattern recognition, supporting non-invasive risk stratification and guiding selection of appropriate surgical or adjuvant therapies.
Core findings summary. TVUS demonstrates high sensitivity (76%-96%) making it an effective initial screening tool for EC in AUB patients, but its specificity (61%-86%) and staging capability are limited. MRI remains the preferred modality for definitive local staging due to superior soft tissue contrast. 3D-US, CEUS, and elastography provide complementary tissue-level information and approach MRI accuracy in selected early-stage cases.
AI-enhanced ultrasound potential. AI and radiomics-based US models demonstrate pooled AUCs up to 0.91 for risk prediction and lesion segmentation, offering promising avenues for personalized EC diagnosis and staging. These models can analyze complex imaging features beyond human perception, potentially enabling earlier identification of subtle pathological changes and molecular subtype prediction without tissue biopsy.
Requirements for clinical translation. Standardization of imaging acquisition protocols, multicenter validation studies, regulatory approval processes, and solutions to operator dependency and equipment variation are needed before widespread AI-enhanced US adoption. Future clinical protocols integrating multi-modal imaging (TVUS, MRI, CT, AI) tailored to institutional resources and patient risk profiles will maximize diagnostic efficiency and improve outcomes in EC.