Endometrial cancer is the most common gynecological cancer in high-income countries. Accurate preoperative assessment matters greatly because it guides surgeons in deciding how extensive an operation to perform. Key factors - like whether the tumor has grown deep into the uterine wall, whether lymph nodes are involved, and what type of tumor it is - all influence treatment planning.
MRI is the standard imaging tool used before surgery, but reading these scans still relies heavily on the skill and experience of the radiologist, and results can vary. Radiomics uses computers to pull out hundreds of detailed, quantitative measurements from those MRI images - measurements invisible to the naked eye. When combined with artificial intelligence, these tools promise a more objective and consistent way to assess cancer before surgery.
One of the most important preoperative questions is whether the tumor has grown more than halfway through the uterine muscle wall - called deep myometrial invasion (DMI). If DMI is present, more extensive surgery and lymph node removal may be needed. Several AI-based radiomics studies have used T2-weighted and contrast-enhanced MRI sequences to predict DMI before the operation.
Results across studies are promising, with many reporting AUC values above 0.80, meaning the model correctly distinguishes deep from superficial invasion most of the time. However, studies use different MRI protocols, different ways of outlining the tumor region, and different AI methods, making it hard to compare results directly. No single validated approach has emerged for routine clinical use yet.
Knowing whether cancer has spread to nearby lymph nodes (LNM) is critical for planning surgery and deciding whether to remove lymph nodes. Radiomics models trained on MRI scans have been used to predict LNM preoperatively. Similarly, lymphovascular space invasion (LVSI) - cancer cells found within blood or lymph vessels - is another high-risk feature that AI models have tried to predict from imaging.
Both LNM and LVSI prediction remain challenging tasks. While several studies report good performance on their own datasets, external validation on data from different hospitals is rare. The small size of many study populations also limits how reliable these predictions are for real-world use.
Endometrial cancers are divided into histological types - endometrioid tumors (lower risk) and non-endometrioid tumors like serous carcinoma (higher risk). Distinguishing these by imaging rather than biopsy could guide treatment decisions faster. AI radiomics models have shown some ability to separate these groups using MRI texture and shape features.
More recently, researchers have linked radiomics features to the molecular classification system used in endometrial cancer - four groups based on genetic alterations (POLE-mutated, mismatch-repair deficient, p53-abnormal, and no specific molecular profile). This emerging field is called radiogenomics. Early studies suggest that specific MRI patterns correlate with molecular subtypes, potentially allowing non-invasive tumor characterization.
Some studies combine radiomics features with clinical information - like age, tumor size, and preoperative biopsy results - into a single risk score or nomogram. These combined models often outperform either approach alone. The goal is to classify patients into low, intermediate, or high-risk groups before surgery, which can guide decisions about lymph node removal and the need for additional treatments.
Integrating multiple data sources is promising but adds complexity. Radiomics features must be carefully selected to avoid overfitting - a situation where a model performs well on training data but fails on new patients. Regularization methods like LASSO regression help reduce the feature set to the most informative ones.
Despite many encouraging studies, no AI-radiomics tool for endometrial cancer MRI has reached routine clinical practice. Several barriers explain this gap. First, there is no standardized way to perform radiomics - different software tools, segmentation methods, and MRI protocols produce different features, making studies hard to reproduce or combine. Second, most studies are small, single-center retrospective analyses, which tend to overestimate performance.
Reproducibility is a specific concern: the same patient's scan, processed by different software, can yield different radiomics values. Without harmonization of imaging protocols and analysis workflows, it is difficult to build a tool that works reliably across hospitals. Independent external validation in diverse patient populations remains rare and is considered a necessary step before clinical adoption.
The field of AI and radiomics for endometrial cancer is rapidly growing but remains largely at the proof-of-concept stage. Most results published so far are exploratory and require prospective validation before any tool can be recommended for routine use. The authors call for multi-center collaborations, standardized imaging protocols, and independent validation datasets as necessary next steps.
Explainability is another important factor. Clinicians are more likely to trust AI tools when the model can indicate which image features drove its prediction. Future systems should include visual outputs - such as highlighted regions in the MRI scan - that let radiologists and gynecological oncologists understand and verify what the AI has detected. The goal is a tool that supports clinical decision-making, not one that replaces expert judgment.