Endometrial cancer is the most common gynecologic cancer in women worldwide, with incidence rising sharply - from 52,600 new US cases in 2014 to 61,880 by 2019. Despite being common, it is not a single disease: different subtypes have dramatically different outcomes and require different treatments.
Pathologists traditionally classify tumors by looking at stained tissue slides under a microscope. The two main histological subtypes are endometrioid (70-80% of cases, generally better prognosis) and serous (about 10% of cases, more aggressive with more frequent metastasis). But visual classification can be difficult, especially for high-grade tumors that blur the line between subtypes.
The Cancer Genome Atlas (TCGA) project introduced a molecular classification system based on DNA analysis that divides endometrial cancer into four subtypes: POLE ultra-mutated, MSI-high (microsatellite instability high), CNV-low (copy number variation low), and CNV-high. The CNV-high subtype has the worst prognosis. Currently, determining molecular subtype requires expensive and time-consuming DNA sequencing - a barrier in many clinical settings.
The researchers developed a customized convolutional neural network (CNN) called Panoptes, named after the all-seeing giant of Greek mythology. A CNN is a type of artificial intelligence that learns to recognize visual patterns in images, similar to how the human brain processes what it sees - but operating on thousands of images simultaneously.
What makes Panoptes novel is its multi-resolution design. Standard H&E slide images (hematoxylin and eosin staining, the routine stain used in pathology) are digitized at very high resolution. Rather than looking at just one magnification, Panoptes simultaneously analyzes tile sets at 2.5x, 5x, and 10x magnification, allowing the model to capture both large-scale tissue architecture and fine cellular-level detail at the same time - mirroring how experienced pathologists review slides.
The model was trained on 496 digitized slides from 456 patients from the TCGA and CPTAC (Clinical Proteomic Tumor Analysis Consortium) datasets. It was then independently tested on slides from New York University hospitals to verify that the learned patterns generalized to real-world clinical samples from a completely different source.
Panoptes achieved a per-patient AUROC of 0.969 for classifying slides as endometrioid versus serous histological subtype (AUROC measures classification accuracy, where 1.0 is perfect and 0.5 is chance). This means the AI model could reliably distinguish between the two main cancer subtypes from the tissue image alone.
For molecular subtyping - predicting the TCGA categories from the tissue image without DNA sequencing - results were equally impressive. The CNV-high subtype (the most aggressive) was predicted with an AUROC of 0.934, and CNV-low with 0.889. These accuracies exceeded baseline CNN models by 5.8% to 23.3%, demonstrating the value of the multi-resolution design.
The model could also predict specific gene mutations from tissue appearance alone. TP53 mutation (a key cancer driver) was predicted with AUROC 0.873, FAT1 with 0.835, and several other clinically relevant mutations including PTEN, FGFR2, MTOR, and CTCF with AUROCs above 0.7. This is particularly significant because knowing mutation status can guide targeted therapy selection.
A key concern with AI models in medicine is whether they are learning genuinely meaningful biological patterns or arbitrary correlations in the training data. To address this, the researchers extracted and visualized the features learned by the model using a technique called t-SNE (t-distributed stochastic neighbor embedding), which reduces complex high-dimensional data to a visual map.
The resulting visualizations showed that the model organized tissue patches in ways that reflected human-interpretable biological features - such as the glandular growth patterns typical of endometrioid cancer versus the papillary patterns of serous cancer. This suggests the AI is genuinely learning pathologically relevant tissue characteristics, not artifacts.
Particularly revealing was the analysis of the CNV-high subtype within endometrioid cases. The model successfully predicted CNV-high status even within what appeared to be the same histological subtype, and the feature analysis confirmed that tumor grade was the key distinguishing factor - high-grade endometrioid tumors that carry CNV-high status look different from lower-grade endometrioid tumors at a cellular level, and the AI detected these subtle differences.
The most demanding test for any AI model is whether it works on completely new data from a different institution - called external validation. The best-performing models were tested on an independent set of slides from NYU hospital patients who were not part of the original training data.
The models retained strong performance on the independent NYU dataset, demonstrating genuine generalizability rather than overfitting to the specific characteristics of the training cohort. This is a critical requirement for clinical deployment - a model must work across different hospitals, staining protocols, and scanner types.
The study also compared Panoptes directly to standard InceptionResnet models (an existing high-performing CNN architecture). Panoptes2 showed significantly higher accuracy than InceptionResnet2 in six prediction tasks at the per-patient level, confirming that the multi-resolution design provides a real performance advantage beyond what could be achieved with a standard single-resolution approach.
Currently, the full molecular classification of endometrial cancer requires DNA sequencing and additional molecular testing, which adds days to weeks of delay and significant cost to the diagnostic process. In many lower-resource settings, this testing is simply not available.
A model like Panoptes could potentially provide molecular subtype information directly from the routine tissue slide that is already prepared for every cancer diagnosis. This would allow faster treatment decisions - particularly for patients with CNV-high disease who may benefit most urgently from aggressive treatment - and could make molecular subtyping accessible in settings that lack sequencing capabilities.
The model's ability to predict specific mutations like TP53 and PTEN from tissue appearance could also serve as a rapid screening tool to flag which patients most need genetic testing, prioritizing sequencing resources toward cases where the results are most likely to change treatment. As pathology increasingly moves toward digital workflows, AI-assisted analysis like Panoptes will become easier to integrate into routine clinical practice.
This study represents a significant step toward a future where a pathologist's standard tissue slide examination yields not just a histological diagnosis but a complete molecular profile - without any additional testing. The Panoptes model demonstrates that meaningful molecular information is encoded in the visual appearance of tumor tissue, even if human pathologists cannot reliably extract it manually.
The 0.969 AUROC for histological subtyping is especially notable because even experienced pathologists sometimes disagree on subtype classification, particularly for high-grade cases. An AI with this level of consistency could improve diagnostic reproducibility across institutions and reduce the inter-observer variability that currently affects treatment decisions.
Future refinements will likely improve performance further, particularly for mutation prediction tasks where some mutations are rare and thus harder to train on. The approach demonstrated here - multi-resolution analysis of standard clinical slides - is applicable to many other cancer types and could transform pathology into a richer, more computationally augmented discipline across oncology.