When a patient is diagnosed with endometrial cancer, one of the most important decisions doctors face is determining the molecular subtype of the tumor. The standard classification system divides endometrial cancer into four subtypes: POLEmut (caused by a mutation in a DNA repair gene), MMRd (mismatch repair-deficient, which means normal DNA error-correction mechanisms have broken down), p53abn (abnormal p53 tumor suppressor protein), and NSMP (no specific molecular profile). These subtypes have very different prognoses and respond differently to treatment.
The problem is that determining molecular subtypes currently requires two expensive, time-consuming laboratory tests: Sanger sequencing to detect gene mutations and immunohistochemistry (IHC) to stain tissue for specific proteins. These tests take days to weeks and require specialized equipment and trained personnel - resources that are limited even in large medical centers, and often unavailable in smaller hospitals in China and other developing regions.
A 2025 study published in Molecular Biomedicine proposed an alternative: instead of running expensive molecular tests, use artificial intelligence to infer the molecular subtype directly from the standard pathology slide - the routine hematoxylin-and-eosin (H and E) stained tissue section that every patient already has. The deep learning pipeline they built achieved over 90% accuracy for all four subtypes and was validated across three separate hospital centers in China.
The AI pipeline consists of three sequential steps. First, low-resolution pathology slide images are sharpened using a tool called SRResGAN - a super-resolution generative adversarial network that increases image quality while preserving microscopic tissue details that would otherwise be blurry. This step is important because diagnostic slides at some hospitals are scanned at lower magnification, and the subtle features needed for molecular subtyping require fine detail.
Second, a transformer-based segmentation model called MedSAM (a medical adaptation of Meta AI's Segment Anything Model) identifies and isolates the actual cancer tissue regions within each slide. Pathology slides contain large areas of non-cancerous tissue, blood vessels, and connective material. By focusing the analysis on the tumor region only, the subsequent classification step receives cleaner, more relevant input. The segmentation model was fine-tuned specifically to identify endometrial cancer lesions.
Third, a deep neural network called ResNet-101 performs the actual molecular subtype prediction. The slide is divided into thousands of small 224x224 pixel patches; each patch gets a subtype probability score, and these scores are aggregated across the full slide to produce a patient-level classification. An additional module using LSTM (Long Short-Term Memory, a type of neural network that handles sequential data well) then combines the predicted molecular subtype with clinical information like age, BMI, and tumor stage to estimate patient survival.
The study enrolled 393 patients diagnosed with endometrial cancer between 2010 and 2018 at Shanghai First Maternity and Infant Hospital. To be included, patients needed complete molecular subtyping confirmed by both Sanger sequencing and IHC, plus at least five years of follow-up data. The molecular subtype distribution was: NSMP 40.7%, p53abn 29.3%, MMRd 18.6%, and POLEmut 11.5% - reflecting the typical distribution seen in Chinese patients, which is somewhat different from Western populations.
The 393 patients provided 15,328 individual histopathological image patches. These were split 80/20 into training and test sets, with care taken to ensure images from the same patient were always in the same set (to prevent artificial inflation of accuracy scores). Five deep learning architectures were compared head-to-head: EfficientNet, DenseNet-161, VGGNet, Vision Transformer (ViT), and ResNet-101. ResNet-101 emerged as the best performer and was selected for final validation.
External validation was conducted at two independent centers: the Obstetrics and Gynecology Hospital of Fudan University (OGHFU, 83 patients) and Pingdingshan Maternal and Child Health Hospital (PMCHH, 35 patients). These centers used different slide scanners and staining protocols, making this a real-world test of whether the model generalizes beyond the single institution where it was developed. External slides were prepared under routine clinical workflows without any special preparation for the study.
ResNet-101 achieved classification accuracies of 92% for both POLEmut and MMRd, 91% for p53abn, and 90% for NSMP. The overall micro-average AUC (a measure that combines performance across all four subtypes) was 0.9838, and the macro-average AUC was 0.9843 - both remarkably high for a classification task with four categories and real-world clinical slides. The POLEmut subtype was classified most accurately, likely because it has distinct microscopic features including increased immune cell infiltration and well-preserved glandular structures that provide strong visual cues.
Interpretability analysis using Grad-CAM (Gradient-weighted Class Activation Mapping) - a technique that highlights which regions of an image most influenced the model's decision - confirmed that the AI was focusing on biologically meaningful areas. For MMRd tumors, the model highlighted glandular regions with dense immune cell infiltration. For p53abn cases, it focused on regions with irregular nuclear shapes and staining patterns. These findings align with what experienced pathologists look for when making similar assessments, suggesting the model has learned genuine diagnostic features rather than spurious patterns.
External validation confirmed strong generalizability. On the OGHFU cohort, AUC was 0.97 with 92% accuracy, 94% specificity, and 87% sensitivity. On the PMCHH cohort, AUC was 0.94 with 93% accuracy, 90% specificity, and 94% sensitivity. These results held despite the different scanners and staining protocols at each center, indicating that the model has learned features that are robust to institution-to-institution variation in slide preparation.
Beyond molecular subtyping, the LSTM survival prediction module demonstrated strong performance at estimating individual patient survival times. The model explained 96.9% of the variance in observed survival durations (R2 = 0.9692), with a mean absolute error of 123 days - meaning the model's survival estimate was typically off by about four months. This level of accuracy in survival prediction is unusual, as most survival models work with categorical groups (high risk vs. low risk) rather than predicting continuous survival time directly.
Kaplan-Meier survival curves based on AI-predicted subtypes closely matched the actual survival curves by true subtype. POLEmut patients showed the most favorable survival trends, consistent with known clinical patterns where this subtype typically has the best prognosis. However, the survival analysis did not reach formal statistical significance (Log-rank P = 0.18), partly because most patients in the cohort were alive at follow-up and events (deaths) were relatively rare. The researchers therefore characterize the survival analysis as exploratory rather than definitive.
Importantly, the model outputs a confidence score for each prediction. Low-confidence predictions - where the model is uncertain between two subtypes - are flagged for confirmatory molecular testing or review by a pathologist. This built-in uncertainty quantification is a clinically important feature: rather than acting as a replacement for all molecular testing, the model is designed to serve as an efficient screening layer, with confirmatory tests reserved for uncertain cases.
The most significant implication of this work is that it could dramatically expand access to molecular subtype-guided cancer care. Current sequencing and IHC-based molecular testing is concentrated in large urban academic centers. Smaller hospitals - which handle the majority of cancer diagnoses, especially in rural and semi-rural China - often cannot access these tests, or face weeks-long delays. An AI model that runs directly on routine pathology slides could deliver preliminary molecular classification within hours of slide scanning, at a fraction of the cost.
This is particularly important for fertility-preserving management in younger patients. Approximately 30% of patients in this cohort were aged 20 to 50, a significant proportion of whom may want to preserve their fertility. Molecular subtype is a key factor in determining whether conservative (non-surgical) management is appropriate. With faster and cheaper molecular profiling, more young women could access personalized treatment planning before opting for hysterectomy.
The study acknowledges important limitations. The retrospective design and predominantly endometrioid tumor composition of the training set mean performance may not fully generalize to rarer histologic subtypes such as serous carcinoma or carcinosarcoma. The external validation cohorts were relatively small (35 and 83 patients). The authors emphasize that model outputs should be treated as preliminary screening rather than standalone diagnostic conclusions, with confirmatory molecular testing for any clinically consequential decisions. Future work will integrate clinical variables and IHC metadata into the survival module and extend validation to larger multi-regional prospective cohorts.