Endometrial cancer - cancer of the lining of the uterus - is the most common gynecological cancer in developed countries. It can be classified into molecular subtypes that help predict how aggressive the cancer is and which treatments will work best.
One of the most clinically important subtypes is microsatellite instability-high (MSI-H). About 30% of endometrial cancers are MSI-H, and these tumors respond particularly well to immunotherapy drugs called checkpoint inhibitors (such as pembrolizumab and nivolumab), which are now recommended by clinical guidelines for advanced MSI-H cases.
Currently, identifying MSI-H requires additional laboratory testing - such as immunohistochemistry or DNA sequencing - which takes extra time and cost beyond the standard tissue biopsy. A faster, cheaper method is needed to make MSI testing accessible for all patients.
This study developed a deep learning framework that can assess MSI status directly from standard H&E (hematoxylin and eosin) stained tissue slides - the same slides already prepared for routine pathological examination. No additional laboratory testing is needed.
The approach is significant because H&E slides are already available for virtually every biopsy case. If AI can reliably read MSI status from these slides, it could provide this important prognostic and treatment-guiding information at no additional cost or delay.
The model was validated on a large, well-characterized dataset of 529 endometrial cancer patients from The Cancer Genome Atlas (TCGA), with confirmed MSI status from molecular testing serving as the ground truth.
Whole slide images are extremely large - often billions of pixels - making it impossible to process them as a single image with standard deep learning methods. The framework addresses this by first selecting informative foreground patches (regions containing tumor tissue) from each slide.
An iterative patch sampling strategy is then used to efficiently sample patches from the selected foreground regions. This ensures the model sees a representative sample of the tumor's spatial heterogeneity rather than focusing on a single region.
The final classification is made using a weighted softmax integrated decision model that combines predictions from multiple patches into a single slide-level MSI prediction. This aggregation approach handles the variability within each slide while being computationally efficient.
The model was trained and evaluated on 529 endometrial cancer whole slide images from the TCGA database, with confirmed MSI/MSS status from molecular testing used as the reference standard. This dataset covers both low-grade (G1, G2) and high-grade (G3) endometrioid carcinoma.
The researchers evaluated performance separately for G1G2 (lower grade, typically less aggressive) and G3 (higher grade, more aggressive) tumors because these subtypes have different tissue appearances that affect how easily MSI markers can be identified visually.
The model was compared against four previously published state-of-the-art methods for MSI prediction from whole slide images, providing a comprehensive benchmark of relative performance.
For low-grade (G1G2) endometrial carcinoma, the model achieved an F-measure of 96%, accuracy of 94%, precision of 93%, and sensitivity of 100%. These are remarkably high results for an automated pathology task of this complexity.
For high-grade (G3) tumors - which are harder to classify because their tissue appearance is more chaotic and less distinctive - the model achieved F-measure of 87% and accuracy of 84%, still significantly outperforming all four benchmark methods.
The model statistically outperformed all four state-of-the-art comparison methods (p < 0.001) for all metrics in both tumor grade groups, demonstrating a substantial and consistent improvement over the prior state of the art.
A critical practical consideration is speed. The framework processes each whole slide image in just 1.03 seconds on a single GPU. This is fast enough to integrate into routine clinical workflows without creating delays in the pathology pipeline.
The combination of high accuracy and fast inference means the system could screen all incoming endometrial cancer slides for MSI status automatically, flagging MSI-H cases for consideration of immunotherapy without requiring the ordering of separate molecular tests.
This could be particularly impactful in healthcare settings where access to molecular testing is limited due to cost or infrastructure, making immunotherapy eligibility assessments more equitable across different patient populations.
This study demonstrates that deep learning can reliably assess MSI status in endometrial cancer directly from routine H&E stained slides, with performance that significantly exceeds previous methods.
The practical implications are significant: a fast, accurate, non-invasive MSI assessment tool could make immunotherapy eligibility testing routine for all endometrial cancer patients, rather than only those who can access specialized molecular testing.
Future directions include prospective clinical validation, extension to other cancer types where MSI status guides treatment, and integration with broader computational pathology platforms that analyze multiple biomarkers simultaneously from the same tissue slides.