Endometrial cancer is the most common gynecologic cancer in developed countries. One of the most critical factors doctors assess before surgery is whether the tumor has grown deep into the uterine muscle wall - a process called myometrial invasion (MI). When the tumor penetrates more than halfway through the muscle (greater than 50%), the cancer is classified as a more advanced stage and patients need more aggressive treatment.
Traditionally, surgeons send a sample to the pathology lab mid-operation (called a frozen section biopsy) to check invasion depth. But this test is far from perfect - it misses the right answer in 10-20% of cases, which can lead to undertreating or overtreating patients. There is a strong need for a reliable, noninvasive way to assess MI before surgery even begins.
This paper is a systematic review and meta-analysis that pooled data from 19 published studies involving 4,373 patients to evaluate how well radiomics-based machine learning - extracting hundreds of mathematical features from MRI scans - can predict myometrial invasion without any tissue sampling.
Researchers searched major medical databases through November 2024 using terms related to radiomics, machine learning, and endometrial cancer. Studies were included if they used radiomics or deep learning applied to MRI images to predict myometrial invasion and reported enough statistical data to calculate sensitivity and specificity.
From an initial pool of thousands of articles, 19 studies met the inclusion criteria, covering 4,373 patients. Each study was assessed for quality using a standardized tool called QUADAS-2. Statistical analysis used a bivariate random-effects model, which accounts for variation between studies. The researchers also separately analyzed conventional machine learning (CML) methods versus deep learning (DL) models to see which performed better.
Meta-regression and subgroup analysis explored whether study design factors - such as imaging field strength, segmentation approach (2D vs 3D), or the number of training samples - affected performance. Publication bias was assessed using Deeks' funnel plot.
Across all 19 studies, radiomics-based machine learning achieved a pooled sensitivity of 0.79 (correctly identifying 79% of invaded cases), specificity of 0.83 (correctly ruling out invasion 83% of the time), and an area under the summary ROC curve (AUSROC) of 0.89. This places the technology in the "good" to "excellent" range for a diagnostic test.
Deep learning models outperformed conventional machine learning: DL achieved sensitivity of 0.81 and specificity of 0.86, versus CML's sensitivity of 0.77 and specificity of 0.81. The difference was statistically meaningful. Notably, 95% of studies used MRI as the imaging modality, confirming MRI as the dominant platform for this application.
However, significant heterogeneity was found across studies - meaning results varied considerably from study to study. This variability appears linked to differences in MRI protocols, patient populations, and how tumor regions were drawn on scans (segmentation methods). No significant publication bias was detected.
Subgroup analyses revealed several important patterns. Studies using 3.0 Tesla MRI (a stronger magnet) tended to perform better than those using 1.5T scanners, likely because higher field strength produces sharper images with more extractable texture details. Studies using both T2-weighted and diffusion-weighted imaging (multi-sequence MRI) generally outperformed those using single sequences.
Segmentation approach also mattered: studies using 3D volumetric tumor delineation showed slightly higher performance than 2D slice-only analysis, suggesting that capturing the full tumor volume provides richer radiomic features. Larger training datasets (more patients used to build the AI model) were also associated with better generalization.
A critical limitation was that 74% of studies were single-center - meaning the AI was trained and tested at the same hospital. Models trained and tested at a single institution typically perform better than they would in practice because they benefit from consistent imaging equipment and patient demographics. Almost no studies included prospective external validation at independent hospitals.
Accurate preoperative knowledge of myometrial invasion depth would allow surgeons to better plan the extent of lymph node removal (lymphadenectomy). Patients without deep invasion who are classified correctly could potentially avoid the complications of extensive node dissection, including lower limb lymphedema - a painful, chronic swelling that affects quality of life.
The meta-analysis results suggest that radiomics-based AI is already approaching clinical utility. With an AUSROC of 0.89, these models are considerably better than intraoperative frozen section analysis in terms of reproducibility. However, they are not yet ready to replace pathology outright - the 79% sensitivity means roughly 1 in 5 deeply invaded cases would still be missed.
For AI to be adopted in routine gynecologic oncology, future studies must demonstrate performance at multiple independent hospitals with different scanners and patient populations. Standardized imaging protocols and fully automated tumor segmentation (removing the need for a radiologist to manually outline the tumor) are also needed before widespread clinical deployment.
This meta-analysis provides the most comprehensive evidence to date that radiomics-based machine learning can detect myometrial invasion with clinically meaningful accuracy. Deep learning models show a consistent edge over traditional handcrafted radiomic methods, likely because neural networks can learn complex, non-linear tumor patterns that manual feature engineering misses.
The authors conclude that while the field shows strong promise, it has significant methodological growing pains. The lack of external validation, small individual study sizes, and inconsistent reporting standards mean that no single model is ready for clinical adoption today. The field needs multicenter prospective trials that follow IBSI (Image Biomarker Standardization Initiative) reporting guidelines.
The consistency across 19 independent studies is encouraging - even with varying methods, AI reliably outperforms chance by a wide margin. As standardization improves and deep learning architectures are pretrained on larger gynecologic imaging datasets, MRI-based AI assessment of myometrial invasion is likely to become a standard part of endometrial cancer workup within the next few years.