Endometrial cancer (EC) is one of the most common cancers affecting women, with an 83% five-year survival rate when caught early. It is most prevalent in developed countries such as those in Europe and North America, driven by sedentary lifestyles, rising obesity, increasing life expectancy, and hormone replacement therapy. Each year, roughly 42,000 women die from the disease, and the vast majority of cases are diagnosed after menopause, with only 4% of patients under the age of 40.
Two biological subtypes have been recognized for decades. Type 1 EC is oestrogen-dependent and generally carries a better prognosis, while Type 2 EC is oestrogen-independent and tends to be more aggressive. Newer classification systems based on genomic and molecular alterations (including biomarkers like ER/PR expression, PTEN expression, DNA mismatch repair loss, and Ki-67/MIB-1) have improved treatment prediction by further subdividing the disease according to its molecular profile.
Diagnostic limitations: Current diagnosis relies on histopathology and staging, supported by imaging modalities such as transvaginal ultrasound (TVUS) and MRI. However, these tools have significant shortcomings. TVUS is operator-dependent and requires a skilled practitioner. MRI is more specific but remains limited to tertiary care settings and can only produce qualitative data. Differentiating between endometrial hyperplasia, atypical hyperplasia, and frank malignancy remains challenging, with disagreements among gynecologic pathologists leading to under- or overtreatment.
The AI opportunity: This 2024 narrative review, published in Annals of Medicine and Surgery, surveys the current landscape of artificial intelligence applications in EC across diagnosis, staging, treatment, and prognosis. The authors argue that AI can address the gaps in traditional diagnostic methods by analyzing large datasets, extracting quantitative imaging features, and detecting patterns that human observers may miss. The review covers applications spanning deep learning, machine learning, convolutional neural networks, MRI-based analysis, hysteroscopy, digital pathology, robotic surgery, and population-level screening.
The review provides an overview of AI technologies relevant to cancer diagnostics. At its core, artificial intelligence in medicine has two components: a virtual side (encompassing machine learning and deep learning) and a physical side (including care robots). Deep learning uses artificial neural networks (ANNs) that mimic the mathematical structure of biological neurons. These networks learn from repetitive modifications to their internal parameters, processing large volumes of data many times before producing an output.
Machine learning (ML) approaches range from simple decision trees to complex deep learning architectures that create patterns via multi-layered neural networks. Computer vision, a subfield of AI, enables machines to identify, analyze, and process medical images and videos in ways that parallel human visual assessment. In medicine, computer vision applications include reading radiographs, MRI scans, mammograms, CT images, and echocardiograms.
The authors note that AI has achieved its greatest success in radiology, where it minimizes workload, reduces errors from untrained readers, and speeds up diagnosis in emergencies. For ultrasound, which is highly dependent on practitioner skill, AI-based positioning and image quality enhancement can reduce the risk of human-driven errors. However, AI-driven reports still require human oversight. The review also highlights multiomics platforms, including repositories such as Oncomine, UALCAN, LinkedOmics, and miRDB, which enable the analysis and visualization of combined genomic, transcriptomic, proteomic, and metabolomic data for endometrial cancer research.
Before exploring AI applications, the review establishes the baseline of how EC is currently diagnosed and treated. Patients typically present with abnormal uterine bleeding, including postmenopausal bleeding, intermenstrual bleeding, or prolonged menstruation. Diagnosis is based on histopathology and staging, with TVUS and MRI used to assess myometrial invasion. Key thresholds include an endometrial thickness greater than 5 mm in postmenopausal women and greater than 15 mm in premenopausal women as indicators warranting further investigation.
Biopsy methods: The gold standard for tissue diagnosis has shifted from dilation and curettage (D&C) to hysteroscopy-guided biopsy, which reduces the risk of malignancy spread. The Pipelle device, invented in 1984, is a widely used alternative that is more affordable and causes less patient discomfort than D&C. However, Pipelle has a notable failure rate for obtaining adequate samples for histological examination. A German study found agreement between Pipelle and D&C diagnoses in 95.5% of cases, but patient- and provider-related factors contribute to sampling failures.
A critical diagnostic gap: One of the most important challenges is differentiating between endometrial hyperplasia (EH), atypical endometrial hyperplasia (AEH), and cancer. Up to 25% of women with AEH may harbor a concurrent well-differentiated adenocarcinoma. AEH is considered a pre-cancerous lesion, and while hysteroscopy has emerged as a predictive tool, its accuracy for distinguishing the specific macroscopic features of cancer remains low, leading to inappropriate treatment decisions.
Standard treatments: The backbone of EC treatment is total abdominal hysterectomy with bilateral salpingo-oophorectomy, with or without lymph node dissection. When myometrial invasion reaches 50% or more of the uterine wall, or when the tumor is grade 2 or 3, radiotherapy is added to reduce recurrence. Chemotherapy regimens, typically cisplatin plus doxorubicin or carboplatin plus paclitaxel, are reserved for extra-pelvic recurrence or advanced disease. For younger patients wishing to preserve fertility, hysteroscopic resection followed by high-dose progestin therapy has shown promise in early-stage disease.
The review highlights several studies demonstrating how AI can improve the diagnostic accuracy of EC using different data sources, from visual images to genetic profiles and radiological scans.
VGGNet-16 for hysteroscopic image analysis (Zhang et al.): This study used 1,851 hysteroscopic images from 454 patients, pre-processed and expanded into a training set of 6,478 images. A tuned VGGNet-16 deep learning model was trained to classify endometrial lesions, then tested on 250 images. The model's diagnoses were compared directly with those of gynecologists, and the study confirmed that deep learning can effectively assess endometrial lesion images obtained via hysteroscopy, providing a potential real-time diagnostic aid during the procedure.
ANN-based gene diagnostic model (Zhao et al.): Using data from the Gene Expression Omnibus (GEO) and The Cancer Genome Atlas (TCGA), researchers identified 14 genes involved in EC and built a diagnostic model using an artificial neural network. The model demonstrated high sensitivity and specificity for detecting early-stage EC, suggesting that gene expression data combined with AI can support molecular-level early diagnosis.
MRI-based deep learning: Multiple studies explored AI with MRI for EC detection. Tao and colleagues demonstrated that deep learning enhances the efficiency of image feature extraction from MRI scans and improves patient classification. A separate study showed that convolutional neural networks (CNNs) applied to MRI achieve high diagnostic performance. A systematic review and meta-analysis on uterine body cancers found that AI-based assessment of radiological, histological, and biochemical modalities can improve treatment effectiveness and prognosis, and that using omics data with AI can surpass traditional classification systems. The review also noted that combining AI with bioimpedance measurement can differentiate between benign and malignant endometrial tumors.
Population-level screening (Hart et al.): This study evaluated 7 different AI algorithms for population-based EC risk assessment and found that random forest performed best. The results indicated that AI-based screening would be both cost-effective and non-invasive compared to traditional screening methods, offering a path toward large-scale early detection.
As pathology laboratories increasingly adopt digital workflows, whole-slide images (WSIs) have become approved for primary diagnostic use. The review examines how AI can leverage this digitization to improve the cytological classification of endometrial tissue samples, particularly in addressing the limitations of Pipelle biopsy specimens.
ANN-based cytological classification (Makris et al.): A study from Greece developed an artificial neural network based on multi-layer perceptron (ANN-MLP) architecture to discriminate between benign and malignant endometrial nuclei and lesions in cytological specimens. The results were strong: for numeric classification, the system achieved 90.87% overall accuracy, 93.03% specificity, and 87.79% sensitivity. When using a percentage-based classifier, performance improved further to 95.91% accuracy, 93.44% specificity, and a remarkable 99.42% sensitivity. These figures demonstrate that computerized systems based on ANNs can reliably aid in cytological classification.
WSI malignancy detection (Fell et al.): A cross-sectional study from Scotland trained an AI model to categorize endometrial biopsy whole-slide images as either "malignant," "other or benign," or "insufficient." The final model accurately classified 90% of all slides correctly and 97% of slides in the malignant class. This suggests that AI-based screening of whole-slide images could meaningfully reduce the workload of pathologists by automatically triaging slides and flagging those most likely to contain cancer.
CT-based skeletal muscle biomarker (Kim et al.): Taking a different approach, Kim and colleagues used AI to analyze CT-based waist skeletal muscle volume as a potential biomarker for EC prognosis. The study suggested that skeletal muscle mass, measured through AI analysis of routine CT scans, could serve as an independent predictor of disease outcomes, adding a novel dimension to prognostic assessment.
The FIGO staging system, developed in 1988 by the International Federation of Gynaecologists and Obstetricians, was revised in 2009 after surveying 42,000 women who underwent surgical staging. The revised system divides EC into detailed subtypes across four stages (I through IV), incorporating myometrial invasion depth, lymphovascular space invasion (LVSI), cervical stromal involvement, and distant metastasis. AI has begun contributing to multiple aspects of staging, treatment planning, and surgical execution.
Robotic-assisted surgery: Several studies evaluated AI-powered robotic systems for EC staging and treatment. Lowe et al. reported on a multi-institutional experience with robotic-assisted hysterectomy showing that robotics may be an excellent intervention for EC staging. Cardenas-Goicoechea et al. compared robotic and traditional laparoscopic surgery, finding no significant differences in 3-year survival (93.3% vs. 93.6%) or disease-free survival. Gocmen et al. demonstrated that robotic-assisted surgery offered advantages including shorter hospital stays, reduced blood loss, and fewer lymph nodes dissected. Bell et al. found that robotic hysterectomy provides better lymph node retrieval than laparotomy and laparoscopy when performed by a skilled surgeon, with faster patient recovery and reduced postoperative morbidity.
Deep learning for lymph node metastasis prediction (Feng et al.): A Chinese study developed a deep learning model to predict lymph node metastasis (LNM) from hematoxylin and eosin (H&E)-stained histopathological images. Trained on various histological subtypes of EC slides, the model created a novel DL-based biomarker that predicted metastatic status with enhanced accuracy, especially for patients in early stages. The model was validated using external data, strengthening confidence in its generalizability.
Radiomics for preoperative assessment (Lecointre et al.): A systematic review from France investigated the contribution of radiomics to radiological preoperative assessment of EC patients. Preliminary data indicated that these technologies, when combined with human intelligence, can address certain clinical problems, although there was not yet enough evidence to fully support routine use in EC treatment. The study also proposed a reproducible AI Quality Score applicable to both machine learning and deep learning studies. Additionally, a clinical calculator (Mysona et al.) was developed to predict the benefit of chemotherapy in stage IA uterine papillary serous cancer (UPSC), finding that low-risk patients would not benefit from chemotherapy.
Beyond diagnosis and staging, AI has been explored for predicting which patients are most likely to develop EC and which are at risk of recurrence after treatment. The review covers several studies that used clinical data and various ML algorithms to build predictive models.
Risk prediction with multiple algorithms (Erdemoglu et al., 2023): This study included 564 patients and tested six different AI algorithms: random forest, logistic regression, multi-layer perceptron, CatBoost, XGBoost, and Naive Bayes. The models used features including age, menopause status, abnormal bleeding, obesity, hypertension, diabetes mellitus, smoking, endometrial thickness, and breast cancer history. Results were promising, with 94% accuracy for predicting pre-cancerous disease and an area under the curve (AUC) of 0.938 on the receiver operating characteristic curve. The study identified age, BMI, and endometrial thickness as the most significant risk factors for developing pre-cancerous and cancerous conditions.
Systematic review of AI in gynecologic cancers (Akazawa et al., 2021): Analyzing studies from January 2010 to December 2020, this review found 71 eligible articles out of 1,632, with 13 focused on endometrial cancer. These studies used imaging data (49%) and value-based clinical data (51%) as AI inputs. The primary prediction targets were definitive diagnosis and prognostic outcomes such as overall survival and lymph node metastasis. A major finding was that 90% of studies included fewer than 1,000 cases, with a median dataset size of only 214 cases, significantly limiting the generalizability and reliability of the models.
Recurrence prediction (Akazawa et al.): A separate study attempted to predict recurrence in early-stage endometrial cancer using machine learning methods based on clinical data. However, due to the small size of the dataset, the machine learning classifiers did not perform as well as anticipated. The authors concluded that while early-stage recurrence prediction is feasible with ML algorithms, much larger datasets are needed to achieve clinically useful accuracy levels.
The review devotes considerable attention to the ethical, legal, and practical barriers that must be overcome before AI can be routinely deployed in EC care. The authors identify multiple categories of concern spanning data security, algorithmic bias, transparency, and implementation challenges.
Data security and the "black box" problem: Machine learning-based AI introduces risks including cybersecurity vulnerabilities, algorithmic inconsistencies, and bias. Many deep learning models function as a "black box" where internal decision-making is opaque to assessors, physicians, and patients. The authors stress that researchers must clarify how model outputs and predictions are generated, and that the inability to test and understand how software might fail remains a fundamental barrier to clinical deployment. Addressing this requires making AI models explainable, auditable, and testable.
Ethical frameworks: At the ethical level, the review advocates for human health as the top-level design principle for trustworthy medical AI. Humans remain the duty bearers because current AI systems do not have legal moral standing. The authors call for improved data quality control, elimination of algorithm bias through increased accountability and traceability, and regulation of the entire AI production process. They also emphasize the importance of international cooperation and communication to assess AI's risks and social implications. A multidimensional approach involving policymakers, developers, healthcare practitioners, and patients is described as essential.
Key practical limitations: Several specific challenges constrain AI progress in EC. The field lacks large, diverse, multi-center datasets. Most existing studies are retrospective, single-institution efforts with fewer than 1,000 cases. External validation datasets are scarce, and cervical cancer has received far more AI research attention than endometrial cancer. Women with limited access to quality healthcare bear the highest mortality burden from EC, and AI-based risk awareness tools could help these women seek earlier treatment, but only if the infrastructure and access barriers can be addressed.
The path forward: The authors conclude that AI holds significant potential to improve EC diagnosis, staging, treatment, and prognosis. Machine learning algorithms can analyze large volumes of data from patient records and imaging studies to create predictive models, assist with preoperative planning, provide intraoperative guidance, and monitor treatment responses. However, data bias must be addressed through proper algorithms trained on unbiased, real-time data. Diverse programming teams and regular algorithm audits are essential. While AI cannot replace clinical judgment, it can meaningfully assist clinicians, particularly in resource-constrained settings where medical expertise is limited. The field remains in its infancy, and radiomics is highlighted as an up-and-coming area that may transform endometrial cancer practices in the future.