Endometrial cancer treatment has been transformed by immunotherapy drugs called checkpoint inhibitors, particularly pembrolizumab, which are most effective in tumors with specific molecular characteristics. These characteristics include mismatch repair deficiency (MMR-D), where the tumor's DNA repair machinery is broken and errors accumulate, and high tumor mutation burden (TMB-H), where the tumor carries an unusually large number of mutations overall.
Currently, identifying MMR-D and TMB-H requires a molecular biopsy: tissue must be removed, processed in a specialized lab, and analyzed using techniques such as immunohistochemistry or next-generation sequencing. While these tests are standard in well-resourced settings, they are expensive, require additional time before treatment can begin, and may not be available at all institutions.
This study explored whether a machine learning analysis of contrast-enhanced computed tomography (CE-CT) images, which are already obtained routinely before endometrial cancer surgery, could predict MMR-D and TMB-H status noninvasively, providing a fast and universally available alternative or supplement to molecular testing.
The study enrolled 150 endometrial cancer patients at Memorial Sloan Kettering Cancer Center who had preoperative CE-CT scans as part of routine staging workup. All patients also underwent tumor sequencing using MSK-IMPACT, a comprehensive genomic profiling panel that analyzes hundreds of cancer-associated genes and provides both mutation status and TMB calculations.
The key methodological challenge was to extract features from the CT images that correlate with molecular biology. The researchers used a radiomics pipeline that automatically extracted hundreds of quantitative features from defined regions of the tumor and surrounding tissue.
For feature selection and model building, they used Recursive Feature Elimination (RFE) combined with a random forest classifier. RFE iteratively removes features that contribute least to the model's predictive accuracy until an optimal subset remains, reducing the risk of overfitting to noise in the training data.
For MMR-D classification, the trained model achieved an Area Under the Receiver Operating Characteristic Curve (AUROC) of 0.78 on both training and test datasets, indicating consistent performance without overfitting. An AUROC of 0.78 means the model correctly ranks a randomly selected MMR-D patient above a randomly selected non-MMR-D patient 78% of the time, which is a meaningful level of discrimination for a noninvasive imaging biomarker.
For TMB-H classification, performance was even higher, with an AUROC of 0.87 on the test set. This is notably strong for an imaging-based prediction of a molecular quantity, and suggests that TMB-H in endometrial cancer may be reflected in tumor imaging characteristics more clearly than in other cancer types.
These performance levels position the model as a potential triage tool: not necessarily replacing molecular testing outright, but identifying which patients are most likely to benefit from expedited molecular workup, or providing guidance in settings where molecular testing is unavailable or delayed.
A notable finding was that the most predictive imaging features were not located within the tumor itself but in the peritumoral rim, the zone of tissue immediately surrounding the tumor boundary. Specifically, Gabor edge features from this region were consistently ranked as the most important variables in the model.
Gabor filters are mathematical tools from signal processing that detect oriented textures and edges at different scales and angles in an image. In the context of CT imaging, they capture subtle variations in tissue density and boundary sharpness that the human eye may not consciously register but that reflect underlying tissue architecture.
The prominence of peritumoral features over intratumoral features is biologically interesting. It suggests that the molecular characteristics of the tumor (specifically MMR-D and TMB-H) are reflected not just in the tumor mass itself but in how the tumor interacts with and infiltrates surrounding tissue, which may relate to differences in immune cell infiltration, tissue remodeling, and invasion patterns between molecularly distinct tumor subtypes.
The clinical significance of this work lies in the potential to use an imaging scan that is already performed as a standard of care to provide molecular information that currently requires separate laboratory testing. Preoperative CE-CT is obtained for virtually all endometrial cancer patients to assess disease extent before surgery, meaning the imaging data needed for this analysis already exists for most patients.
If validated prospectively, a radiomics tool like this could be used to flag high-probability MMR-D or TMB-H cases at the time of initial imaging workup, allowing oncology teams to prioritize or expedite molecular testing, or to counsel patients earlier about immunotherapy eligibility as part of their treatment planning.
The authors appropriately note that the model was developed and tested at a single institution, and that external validation at other centers with different patient populations, imaging protocols, and CT scanner types is necessary before clinical deployment. However, the alignment between the machine learning-identified imaging features and known biological differences between MMR-D and non-MMR-D tumors provides biological plausibility that strengthens confidence in the approach beyond the numerical performance metrics alone.