Endometrial cancer (cancer of the uterine lining) comes in several different subtypes that vary widely in how dangerous they are. Correctly identifying the subtype is critical because it determines treatment intensity, but doing so manually requires expert pathologists examining tissue under a microscope, which is time-consuming and subjective.
This study describes EC-AIHIS, an artificial intelligence system built to automatically classify endometrial cancer histotypes (tissue-level subtypes) from digitized microscope slide images stained with hematoxylin and eosin (H&E), the standard stain used in pathology labs worldwide.
The classification follows the FIGO 2023 staging system, which groups endometrial cancer into non-aggressive histotypes (such as endometrioid grade 1-2 and mucinous carcinoma) and aggressive histotypes (such as serous, clear cell, endometrioid grade 3, and carcinosarcoma). This distinction directly affects whether a patient will need more aggressive post-surgical treatment.
EC-AIHIS uses ResNet18, a convolutional neural network architecture widely used in medical image analysis. The model was trained on tissue images from 1,187 patients across multiple hospitals, with 1,067 cases used for internal testing and 68 cases used for external validation at an independent institution.
To assess real-world robustness, the team also tested the model on slides scanned with different scanner types than those used in training, and on slides of deliberately lower image quality to simulate challenging clinical conditions. This kind of robustness testing is essential for understanding how an AI model will perform when deployed outside the controlled environment where it was developed.
The model was evaluated using the Area Under the Receiver Operating Characteristic Curve (AUC), a standard metric where 1.0 is perfect and 0.5 is no better than random guessing. On internal test data the AUC was 0.911, on external validation 0.859, on different scanner types 0.925, and on poor-quality slides 0.818, indicating consistently strong performance across conditions.
To determine whether the AI adds clinical value beyond what a trained doctor can do alone, the researchers compared EC-AIHIS to human pathologists on a challenging subset of cases that showed adenoid structures, a histological pattern that makes subtype classification particularly difficult.
On these ambiguous cases, EC-AIHIS achieved 88% accuracy, while junior pathologists (doctors still in training) achieved 81% accuracy on their own. More importantly, when junior pathologists were given EC-AIHIS output as a reference tool, their accuracy improved substantially from roughly 79% to 94%.
This finding illustrates one of the most valuable potential roles for AI in pathology: not necessarily replacing experts, but serving as a high-quality second opinion that helps less experienced clinicians match the performance of experts. In hospitals and clinics where highly specialized pathologists are not always available, tools like EC-AIHIS could meaningfully reduce diagnostic errors.
Modern endometrial cancer management increasingly relies on molecular subtyping, which classifies tumors based on genetic markers into four groups: POLE-mutant, mismatch repair deficient (MMR-D), copy-number low, and copy-number high (also called the p53-abnormal or p53abn subtype). These molecular subtypes have different prognoses and respond differently to therapies like immunotherapy and chemotherapy.
The researchers found that EC-AIHIS histotype classifications correlated with molecular subtype assignments at an AUC of 0.894, suggesting the visual patterns it detects in tissue images reflect underlying molecular biology.
Crucially, within patients classified into the high-risk p53abn subtype, the AI's histotype predictions were able to further stratify survival outcomes (p=0.048 in a Kaplan-Meier analysis). This means the AI adds prognostic information even within already-high-risk groups, which could help oncologists decide how aggressively to treat individual patients.
Accurate histotype classification is not just an academic exercise: it directly influences surgical decisions (such as lymph node dissection extent), adjuvant chemotherapy or radiation choices, and eligibility for clinical trials. Misclassification can lead to overtreatment or undertreatment, both of which harm patients.
EC-AIHIS addresses a practical gap: H&E staining is performed universally in all pathology labs, whereas molecular testing is expensive, requires specialized equipment, and is not available in lower-resource settings. An AI model that works on standard H&E slides can therefore function as a widely deployable screening tool.
The authors emphasize that EC-AIHIS is designed as an assistive tool rather than a replacement for pathologists. Its best use case is helping less experienced clinicians perform closer to expert level, flagging ambiguous cases for senior review, and providing consistent quantitative outputs that reduce inter-observer variability in diagnosis.