Endometrial cancer diagnosis has traditionally relied on histopathology - a pathologist examining tumor tissue under a microscope and classifying it by how the cells look (morphology). While this remains fundamental, the field is undergoing a transformation as molecular testing reveals that tumors with similar appearances can behave very differently, and vice versa.
The 2023 WHO classification of endometrial cancer integrated molecular subtyping into formal diagnostic criteria, creating four molecularly defined classes: POLEmut (POLE-mutated), MMRd (mismatch repair deficient), p53abn (p53-abnormal), and NSMP (no specific molecular profile). Each class has a characteristic clinical outcome and, importantly, characteristic microscopic features that pathologists can learn to recognize.
This review examines how morphology and molecular classification intersect, what microscopic patterns are associated with each molecular subtype, and how deep learning algorithms trained on digitized tissue slides are emerging as tools to automate molecular subtype prediction - bringing the power of AI to routine pathology diagnosis.
POLEmut (POLE-mutated) EC carries mutations in the DNA proofreading enzyme POLE, generating an extremely high number of mutations. Despite appearing histologically aggressive, POLEmut EC has excellent prognosis - likely because the massive mutation burden makes it highly visible to the immune system. These patients may not need adjuvant treatment even with high-grade features.
MMRd (mismatch repair deficient) EC, also called microsatellite instability-high (MSI-H), results from failure of the DNA mismatch repair system. These tumors also accumulate many mutations and respond well to immune checkpoint inhibitors. MMRd is the most clinically actionable subtype given the availability of FDA-approved pembrolizumab for this group.
p53abn EC carries mutations in the tumor suppressor TP53, often alongside widespread chromosomal instability. This subtype has the worst prognosis and typically corresponds to serous carcinoma histology. NSMP EC lacks defining molecular alterations from the first three categories and has intermediate prognosis - it is the largest and most heterogeneous group, representing a catch-all category that may be further subdivided in the future.
POLEmut EC often displays features that would otherwise suggest poor prognosis: solid growth patterns, bizarre multinucleated tumor giant cells, and marked nuclear pleomorphism. Paradoxically, these tumors are characterized by a dense lymphocytic infiltrate and prominent tumor-infiltrating lymphocytes (TILs), reflecting the active immune response to the hypermutated tumor. Tertiary lymphoid structures (organized immune aggregates) may also be present.
MMRd EC shares some features with POLEmut - including lymphocytic infiltration and TILs - but tends to be less extreme. The pattern is sometimes described as peritumoral lymphocytic infiltration with band-like distribution. p53abn EC is often recognized by papillary architecture, a ragged luminal surface where cells appear to be falling off the gland edges, and brisk (numerous) mitotic figures reflecting rapid cell division.
NSMP EC frequently shows well-formed glandular structures with smooth luminal borders - a more 'well-differentiated' appearance. A characteristic pattern is MELF invasion (microcystic, elongated, and fragmented glands) in the myometrium, and squamous differentiation (areas resembling squamous cells) is common. These morphological associations can help pathologists suspect the molecular subtype even before molecular testing results are available.
The digitization of pathology slides into whole slide images (WSIs) has enabled application of deep learning - the same technology powering facial recognition and autonomous vehicles - to tissue analysis. Convolutional neural networks can be trained to recognize patterns in tissue images associated with specific molecular features, effectively learning to 'see' molecular biology through morphology.
Multiple studies have applied this approach to EC molecular subtyping. Wang et al. reported an AUC of 0.73 for MSI prediction from H&E slides, while Hong et al. achieved up to AUC 0.87 for predicting copy-number high (p53abn) tumors. These results suggest that while deep learning cannot yet replace molecular testing, it can provide a meaningful first-pass estimate from slides that are already routinely collected.
A key challenge is that deep learning models require large training datasets and are prone to learning spurious patterns specific to one hospital's slides (staining protocols, scanner types, patient demographics). External validation - testing models on slides from institutions not involved in training - is essential to demonstrate that models have learned generalizable biological patterns rather than site-specific artifacts.
The integration of morphology and molecular testing changes how pathologists approach EC diagnosis. Rather than first classifying by histological type (endometrioid vs serous vs clear cell) and then performing molecular testing, the new paradigm runs morphology and molecular testing in parallel, with each informing interpretation of the other.
Practically, recognizing morphological features suggestive of a specific molecular subtype allows pathologists to prioritize molecular tests and flag discordant cases for review. A tumor with classical p53abn morphology but an MMRd molecular result would be flagged as a potential double-classifiable tumor - an important edge case with specific treatment implications.
Deep learning tools, once validated, could serve as a quality control layer - alerting pathologists when the morphological and molecular classifications appear discordant, or providing automated screening to identify cases most urgently needing molecular testing in resource-limited settings. This is particularly relevant for global health equity, where molecular testing infrastructure may be limited.
The field of computational pathology - using computers to analyze tissue images - is moving rapidly from research to clinical deployment. Several AI-based pathology tools have received FDA clearance, and major cancer centers are investing in digital pathology infrastructure. EC molecular subtyping represents an ideal application area given the clear clinical implications of each subtype.
Beyond molecular subtyping, future deep learning models may predict specific gene mutations, treatment response, and recurrence risk directly from tissue images. The morphological features encoded in H&E slides reflect thousands of molecular processes - deep learning may be able to decode much of this information systematically, going beyond what trained human pathologists can perceive.
The most important near-term development will be large-scale prospective studies demonstrating that AI-assisted pathology improves patient outcomes, not just diagnostic accuracy metrics. Ultimately, the value of these tools will be measured by whether patients treated using AI-informed diagnoses have better survival, fewer unnecessary treatments, and more appropriate therapies than those diagnosed by conventional pathology alone.