Mismatch repair (MMR) deficiency is one of the four molecular subtypes of endometrial cancer, affecting approximately 30% of patients. When the MMR system - which corrects DNA copying errors - is broken, tumors accumulate thousands of mutations that make them highly recognizable to the immune system. This makes MMR-deficient (MMRd) tumors dramatically responsive to immune checkpoint inhibitor (ICI) therapy.
Currently, MMR status is determined by immunohistochemistry (IHC) - testing for loss of MMR proteins (MLH1, MSH2, MSH6, PMS2) - or by molecular testing for microsatellite instability (MSI-H). IHC costs roughly 1,550 yuan and takes 2 days; next-generation sequencing (NGS) costs 6,000 yuan and takes 15 days. Both add time and expense to the diagnostic pathway.
MMRNet is a deep learning model designed to predict MMR deficiency directly from routine H&E-stained whole-slide images - the standard slides already created during pathological examination. If accurate, this would eliminate the need for additional IHC or NGS testing, reducing cost to approximately 60 yuan per case and delivering results in 1 day.
MMRNet uses an ensemble approach - combining multiple independently trained ResNet18 convolutional neural networks and aggregating their predictions. ResNet18 is a well-established deep learning architecture that processes image tiles extracted from whole-slide images. The ensemble reduces variance compared to any single model, improving consistency across diverse cases.
Each slide is broken into thousands of small tiles, and each tile receives a score reflecting MMRd likelihood. A slide-level aggregation step combines tile scores into a final prediction for the whole slide. This approach handles the challenge that MMRd features may not be uniformly distributed across the entire tumor - the model learns which regions are most informative.
The model was trained on an internal dataset and validated on three independent external cohorts: the publicly available TCGA dataset, a MultiCenter cohort, and the GWCH (Guangzhou Women and Children Hospital) cohort. Using multiple geographically and institutionally diverse validation sets is essential for confirming that the model generalizes beyond its training environment.
On the internal validation cohort, MMRNet achieved an AUROC of 0.897, with sensitivity 0.628 and specificity 0.949. The high specificity means the model rarely misclassifies MMR-proficient tumors as deficient (few false positives), but the moderate sensitivity means it misses some genuine MMRd cases. This reflects the conservative calibration appropriate for a screening tool where false positives could expose patients to unnecessary immunotherapy.
On the three external validation cohorts, MMRNet achieved AUROCs of 0.790 (TCGA), 0.807 (MultiCenter), and 0.863 (GWCH). The consistency across these independent datasets - including publicly available TCGA data and institutional cohorts - supports the model's generalizability. The GWCH cohort, collected under most similar conditions to training, showed the strongest performance.
A particularly important finding was that MMRNet outperformed all pathologist experience levels on sensitivity. When pathologists were asked to assess MMR status from H&E slides alone (without IHC), experienced pathologists achieved lower sensitivity than MMRNet, confirming that MMRd features are subtle enough that current human practice cannot reliably detect them without molecular testing.
The study tested human-machine fusion - combining MMRNet predictions with pathologist assessments to see if their combination outperforms either alone. Multiple fusion strategies were tested, including weighted averaging of scores and sequential screening approaches.
The best fusion strategy - Expert Pathologist combined with MMRNet - achieved an AUROC up to 0.9322, higher than either the expert pathologist or MMRNet alone. This demonstrates complementarity: the AI captures patterns the pathologist misses, while the pathologist catches cases where the AI prediction is uncertain.
The fusion approach is practically relevant because it positions MMRNet as a collaborative tool rather than a replacement for pathologists. In a workflow where the AI flags uncertain cases for expert review, the combination maximizes accuracy while preserving appropriate human oversight for ambiguous or borderline predictions.
The cost analysis is compelling: MMRNet costs approximately 60 yuan per case in 1 day, compared to IHC at 1,550 yuan in 2 days and NGS at 6,000 yuan in 15 days. In high-volume cancer centers processing hundreds of EC cases monthly, the savings across both cost and turnaround time would be substantial.
For treatment decision timelines, the speed advantage is clinically meaningful. Patients awaiting MMR status results to determine eligibility for immunotherapy face delays that affect both treatment planning and patient anxiety. A same-day AI result from existing slides accelerates the entire pathway from diagnosis to treatment initiation.
A phased implementation strategy is proposed: MMRNet first as a pre-screening tool that flags high-confidence MMRd cases (avoiding IHC for those), with IHC reserved for borderline or high-confidence MMR-proficient cases where the clinical stakes are highest. Prospective clinical validation comparing this strategy to current standard care will be needed before regulatory approval.