When a surgeon removes an endometrial cancer tumor, a pathologist creates H&E (hematoxylin and eosin) stained slides from the tissue. These slides reveal the shape and arrangement of cancer cells under the microscope. This paper introduces HECTOR, a deep learning model that analyzes these slide images alongside clinical stage information to predict whether a patient's cancer will return after treatment.
The goal is to identify patients at high risk of distant recurrence (cancer returning in organs far from the uterus, like the lungs or liver), who may need more aggressive follow-up or additional therapy beyond standard treatment.
HECTOR is a multimodal model, meaning it combines information from two different sources: the visual patterns in whole-slide H&E images and the patient's anatomical tumor stage. Whole-slide images (WSIs) are very high-resolution digital scans of entire tissue sections, containing billions of pixels of information that no human could efficiently review in detail.
The model also incorporates the tumor's molecular classification (based on genetic markers like POLE mutations or mismatch repair status), which is increasingly used to stratify endometrial cancer patients. Deep learning allows HECTOR to automatically discover subtle visual patterns associated with recurrence risk that may not be visible to the human eye.
HECTOR was evaluated using the C-index, a measure of how well the model discriminates between patients who will or will not recur. A C-index of 1.0 is perfect discrimination; 0.5 means no better than chance. HECTOR achieved a C-index of 0.789 in the internal validation set and 0.828 and 0.815 in two independent external hospital datasets (UMCG and LUMC), demonstrating consistent performance across different patient populations and institutions.
This external validation is critically important because it shows the model works beyond the specific hospital where it was developed.
HECTOR stratified patients into three risk groups: low, intermediate, and high. The 10-year distant recurrence-free survival probabilities were strikingly different across these groups: approximately 97.0% for the low-risk group, 77.7% for the intermediate-risk group, and 58.1% for the high-risk group.
These large separations between groups mean HECTOR can meaningfully distinguish patients who are almost certainly cured from those with a substantial chance of relapse, providing actionable clinical information.
Currently, many endometrial cancer patients receive similar follow-up schedules regardless of their individual recurrence risk. HECTOR's ability to identify a low-risk group with 97% 10-year recurrence-free survival means those patients might safely receive less intensive surveillance, reducing the burden of frequent clinical visits and tests.
Conversely, the high-risk group (58% recurrence-free at 10 years) could be prioritized for clinical trials of adjuvant therapy (additional treatment after surgery) aimed at preventing relapse. The model uses information already collected during routine pathology processing, so no additional tests or costs would be required.
Deep learning models for pathology work by breaking each gigapixel whole-slide image into thousands of smaller tiles. The model learns which tile-level patterns (cell shapes, tissue architecture, immune cell infiltration) are most associated with the outcome of interest. An attention mechanism then weights the most informative regions, allowing the model to focus on the parts of the slide that matter most.
HECTOR's use of both visual and clinical features means it captures complementary information: the slide reveals microscopic biology while the stage and molecular classification provide macro-level context. This combination outperforms either data type used alone.
HECTOR represents a significant step toward AI-assisted pathology in endometrial cancer care. By analyzing the same tissue slides already prepared during routine pathology workup, the model adds prognostic information at essentially no additional cost or patient inconvenience.
The model's consistent performance across multiple independent European hospital datasets suggests it is generalizable rather than over-fitted to a single institution's patient population. Future work should focus on prospective validation and integration into clinical decision support systems to guide adjuvant treatment decisions.