Modern endometrial cancer treatment increasingly depends on molecular classification rather than histological type alone. The Cancer Genome Atlas (TCGA) and the related ProMisE system classify endometrial cancers into four molecular subtypes: POLE-mutated, Mismatch Repair Deficient (MMRd), p53-abnormal, and No Specific Molecular Profile. Each subtype predicts different outcomes and guides different treatment choices.
Mismatch repair (MMR) deficiency is one of the most clinically important of these classifications. MMR proteins (MLH1, MSH2, MSH6, PMS2) normally fix DNA copying errors. When these proteins are absent or non-functional - a condition called dMMR - errors accumulate rapidly throughout the genome, causing microsatellite instability (MSI). Endometrial cancer has the highest MSI frequency of any solid tumor at approximately 17%, making MMR status uniquely important in this cancer type.
dMMR status predicts response to immune checkpoint inhibitors (ICIs) - drugs like pembrolizumab and durvalumab that re-activate the immune system to attack cancer. Clinical trials such as PORTEC-RAINBO are specifically investigating immunotherapy for dMMR endometrial cancer patients. However, current MMR testing by immunohistochemistry (IHC) or PCR requires additional laboratory steps, time, and cost after initial biopsy - creating a barrier to universal molecular profiling.
The study used archived tissue from 114 endometrioid endometrial cancer patients treated at Sapporo Medical University Hospital (2005-2009). For each patient, the MMR status was established by immunohistochemical staining of the tumor tissue for two MMR proteins: PMS2 and MSH6. This yielded 29 dMMR (deficient) and 85 pMMR (proficient) cases - a 1:3 imbalance that would later challenge model training.
Each H&E-stained glass slide was digitized using a Nanozoomer whole-slide scanner, creating high-resolution digital files. Each large whole-slide image was then divided into non-overlapping 512x512 pixel tiles at 5x magnification, yielding an initial 97,547 tiles. Tiles dominated by non-tumor tissue (stroma, artifacts, poor preparation) were systematically excluded, leaving 5,397 high-quality tumor-rich tiles - just 5.5% of the original total, highlighting how much rigorous quality filtering matters.
A critical data imbalance problem was addressed through downsampling: since the pMMR class had roughly 2.6 times as many tiles as dMMR, training without adjustment caused the model to predict pMMR for nearly everything. Downsampling pMMR tiles to match dMMR count dramatically improved recall from 0.09 to 0.88, demonstrating that class balance is as important as model architecture for effective training on imbalanced pathology datasets.
Six convolutional neural network (CNN) architectures were tested. ResNet50 performed best among all models, achieving the highest AUROC of 0.91 with 80% accuracy, 76% precision, 88% recall, and 0.81 F-score on the test dataset. ResNet101 performed nearly identically (AUROC 0.89). All CNN architectures achieved AUROC above 0.8, suggesting that convolutional features are well-suited for this task.
Two types of Vision Transformer (ViT) models - which use attention mechanisms instead of convolution to process images - significantly underperformed: ViT_B16 achieved only AUROC 0.62 and ViT_B32 reached 0.76. The authors attribute this to ViTs' lack of built-in convolutional structure, suggesting that local spatial features - which convolution specifically captures - are particularly important for histopathology classification tasks.
A custom API-Net-based model combining an attention mechanism with a ResNet backbone achieved AUROC 0.89, matching ResNet50's performance while adding the interpretability benefit of attention maps. This model processes pairs of images simultaneously and learns representative prototypes for each MMR class, then classifies new images by similarity to these prototypes - a concept related to few-shot learning metric approaches.
The study's core clinical proposal is to use deep learning-based MMR prediction as a cost-effective pre-screening step before expensive molecular testing. In a proposed workflow, AI would first analyze H&E slides (which are already produced for routine diagnosis) to classify cases as likely dMMR or pMMR. Only cases requiring confirmation would then undergo IHC or sequencing, potentially reducing the number of costly molecular tests by a significant fraction.
Cost-effectiveness analyses in colorectal cancer have estimated that AI-assisted molecular profiling could reduce testing costs by approximately 12.9% while maintaining high diagnostic accuracy. The authors propose that similar savings could apply to endometrial cancer - where universal molecular profiling is increasingly recommended but not yet financially feasible in all healthcare settings.
The clinical urgency is increasing: the PORTEC-RAINBO trial's MMRd-GREEN arm is actively comparing immunotherapy plus radiotherapy versus radiotherapy alone specifically for dMMR endometrial cancer. As immunotherapy trials like this mature and gain regulatory approval, accurate and accessible dMMR identification will become a standard step in endometrial cancer care - making AI-assisted pre-screening highly relevant.
Convolutional neural networks (CNNs) process images by applying learned filters that detect local patterns at progressively larger scales - starting with edges and textures, building up to complex structures like gland shapes and nuclear arrangements. This inductive bias toward local spatial features is well-matched to histopathology, where diagnostic information often appears in the local microarchitecture of tissue.
Vision Transformers (ViTs) use a different mechanism called self-attention, which allows them to relate any part of the image to any other part simultaneously. While ViTs excel at many image recognition tasks when trained on very large datasets, they lack the local feature emphasis of convolutions. In this small-dataset pathology setting (5,397 tiles), pure ViTs failed to learn useful representations, suggesting they may require far more data to outperform CNNs in histopathology tasks.
The authors' custom API-Net-based model (AUROC 0.89) offers a promising hybrid direction: it incorporates ResNet as its backbone (providing convolution's local feature strengths) while adding an attention mechanism to highlight which image regions are most informative for MMR classification. This hybrid approach may represent the most productive direction for future improvements over baseline CNN performance.