Endometrial cancer (EC) is the most common gynecologic cancer in the US, and pathologists must assign a tumor grade - low or high - to guide treatment decisions. This grading is currently done by examining tissue under a microscope, a process that is time-consuming and subject to inter-observer variability between pathologists.
The challenge is especially pronounced for grade 2 tumors, which sit in the middle of the spectrum and are notoriously difficult to classify consistently. Disagreements between pathologists on grade 2 cases can directly affect treatment planning and patient outcomes.
Deep learning models trained on whole-slide images (WSIs) - high-resolution digital scans of tissue slides - offer a promising path to faster, more consistent grading. This study developed and tested EndoNet, a model specifically designed to classify EC as low-grade (grades 1 and 2) or high-grade (grade 3).
The researchers developed EndoNet by combining two types of neural networks: a convolutional neural network (CNN) to identify local tissue patterns in small image tiles, and a vision transformer (inspired by the MaskHIT architecture) to integrate those tile-level patterns across an entire slide.
Training used 929 whole-slide images from patients treated at Dartmouth Health. Each WSI was divided into small image tiles, and the model learned which tile patterns are most predictive of tumor grade. The transformer layer then combined evidence from all tiles to make a final slide-level prediction.
The model was tested on two independent sets: an internal test set of 110 patients from Dartmouth Health, and an external validation set of 100 patients from The Cancer Genome Atlas (TCGA), a publicly available national cancer database. Using an external dataset is critical for confirming that the model generalizes beyond the institution where it was built.
On the internal test set, EndoNet achieved an F1 score of 0.91 and an area under the curve (AUC) of 0.95, indicating very strong ability to correctly distinguish low-grade from high-grade tumors. F1 score combines precision and recall into a single metric, and AUC measures the model's overall discriminative power.
On the external TCGA validation set, performance remained strong with an F1 of 0.86 and an AUC of 0.86. This consistency across datasets from different institutions suggests that EndoNet captures genuine biological differences rather than institution-specific quirks in slide preparation.
Grade 2 tumors were the hardest to classify, with accuracy of 0.81 on the internal test and only 0.68 on the external TCGA set. This mirrors the difficulty human pathologists face with grade 2 cases and highlights a key remaining challenge for AI-based grading systems.
The lower accuracy on grade 2 tumors is not surprising: these tumors are biologically heterogeneous and contain a mix of glandular and solid tissue patterns that overlap with both grades 1 and 3. Even experienced pathologists show significant disagreement on grade 2 classification, so the model is reflecting a genuine biological ambiguity.
The researchers also examined which tissue regions EndoNet focused on when making predictions, using attention map visualizations. The model's attention aligned reasonably well with features that pathologists use, such as areas of solid tumor architecture. This interpretability is important for clinical trust.
A key limitation is that the model was trained to reproduce pathologist labels, which themselves have known inconsistency. Future work could explore using molecular subtype information - which provides a more objective measure of tumor biology - to generate more reliable training labels for grade 2 cases.
EndoNet is designed as a decision-support tool rather than a replacement for pathologists. In clinical practice, it could flag cases where the AI grade disagrees with a preliminary pathologist read, prompting a second review - particularly useful for the grade 2 cases where human agreement is lowest.
A major practical advantage is throughput: once a slide is digitized, EndoNet can analyze it in seconds. At large cancer centers processing hundreds of biopsies per week, automated pre-screening could meaningfully reduce pathologist burden and turnaround time.
The external validation on TCGA data demonstrates the model's potential generalizability, but prospective clinical trials - where the model grades slides in real time before pathologist review - will be needed before deployment in clinical settings. The researchers emphasize that site-specific fine-tuning may be necessary as slide preparation protocols vary between institutions.