AI Reads Tumor Slides and Genes Together to Flag Hidden Risk in Early-Stage Endometrial Cancer

For most women diagnosed with early-stage endometrial cancer, surgery is curative, and the outlook is reassuringly good. Yet a stubborn minority of patients with disease that appears confined to the uterus will go on to develop a recurrence that is typically incurable. An international research team now reports a new artificial intelligence framework, called PRISM, that combines the two richest sources of information available at diagnosis — the ordinary microscope slide and the tumor’s transcriptome — to identify those hidden high-risk cases more accurately than current clinicopathological and molecular tools allow.

The study, published in Frontiers in Cell and Developmental Biology, addresses a persistent gap in gynecological oncology. Endometrial carcinoma is the most common gynecological malignancy in high-income countries, and roughly eighty percent of patients are diagnosed with FIGO stage I–II disease. About eighty-five percent of these tumors are the endometrioid subtype. Current risk assessment relies on stage, histological grade, depth of myometrial invasion, lymphovascular space invasion, and molecular subtype assignment into POLE-mutant, microsatellite-instability, copy-number-low, and copy-number-high classes. Even with molecular classification, which has measurably improved risk assignment, clinically relevant uncertainty remains. Histological grading suffers from interobserver variability, lymphovascular space invasion is difficult to distinguish from morphological mimics, and molecular subtypes must still be interpreted alongside tumor architecture, spatial heterogeneity, and clinical context.

The new framework, formally named the prototype-resolved interpretable survival model, was built to integrate routine haematoxylin and eosin-stained whole-slide images with RNA-sequencing-derived pathway programmes for progression-free survival prediction. The technical design is distinctive. Diagnostic slides are cut into 256-by-256-pixel patches at an equivalent twentyfold magnification, and a pathology encoder converts each patch into a numerical embedding, producing five hundred image tokens per patient. On the molecular side, gene expression is filtered, log-transformed, standardized, and organized into 236 pathway-level tokens drawn from Hallmark and KEGG gene sets. The critical innovation is an adaptive prototype-learning bottleneck that compresses the hundreds of image tokens into just four learnable morphological prototypes — shared tissue states that recur across patients — before any cross-modal fusion occurs.

This compression matters because directly modelling the interactions between hundreds of image patches and hundreds of molecular pathways creates an enormous parameter space that is difficult to estimate reliably in disease-specific cohorts with limited matched data. By first distilling tissue appearance into four prototypes and then aligning those prototypes with pathway-level representations, PRISM reduces the multimodal fusion problem to a manageable size while retaining traceability back to specific regions of the slide. A pathway–prototype alignment module links transcriptional programmes to morphological states, and the fused representation feeds a Cox proportional-hazards output head that produces a continuous risk score, with higher scores indicating greater probability of progression, recurrence, or death.

The development process was deliberately rigorous. The model was trained on 229 patients from The Cancer Genome Atlas uterine carcinoma cohort, with a held-out internal validation set of 58 patients. Only after the architecture, hyperparameters, preprocessing workflow, and checkpoint-selection rule were completely locked was the model retrained on all 287 TCGA cases and evaluated exactly once, without any tuning, in an independent external cohort of 63 patients from the Clinical Proteomic Tumor Analysis Consortium. The CPTAC data played no role whatsoever in training, feature selection, or threshold optimization — a design feature that lends unusual credibility to the external results in a field where leakage between discovery and validation sets is a common criticism.

Across 350 total patients, PRISM achieved Harrell’s concordance indices of 0.770 in the TCGA training cohort, 0.751 in internal validation, and 0.745 in the external CPTAC cohort, indicating that risk ranking held up when transported to a different patient population, sequencing platform, and slide-scanning environment. At the thirty-six-month horizon, the censoring-adjusted time-dependent area under the curve reached 0.727 externally, compared with 0.685 for a baseline model combining clinicopathological variables and molecular subtype. In a complementary fixed-horizon analysis restricted to patients whose thirty-six-month event status was ascertainable, adding PRISM to the clinicopathological model raised the AUC in the external cohort from 0.609 to 0.719. In the TCGA evaluable subset, the corresponding jump was even larger, from 0.697 to 0.905 when molecular subtype was included. Ablation experiments reinforced the architecture’s logic: a version without the prototype bottleneck and an image-only model both performed worse and varied more across five independently generated data splits, while the image-only model collapsed to a concordance index of 0.584 in external validation.

Perhaps the most striking feature of the study is its interpretability. Because the model’s risk predictions pass through named morphological prototypes, a gynecological pathologist could review the patches assigned to each prototype and attach descriptive labels. Prototype one corresponds predominantly to stromal and fibromuscular tissue, prototype two to tumour epithelial and glandular morphology, and prototype three to hypercellular, pleomorphic tumour regions. A fourth prototype captured low-content and artefact-associated tissue edges and was excluded from biological interpretation. Spatial activation maps showed that the first three prototypes occupied distinct regions within individual slides, meaning the model’s risk estimate could be traced to recognizable tissue compartments. High-risk tumors showed lower stromal prototype burden and higher hypercellular prototype burden — a shift from supportive tissue architecture toward aggressive cellular morphology.

The molecular side of the interpretation proved equally coherent. Tumours classified as high risk by PRISM showed elevated pathway-programme scores for proliferation, cell-cycle activity, glycolysis, hypoxia, focal adhesion, and epithelial–mesenchymal transition, while low-risk tumors scored higher on metabolism- and homeostasis-associated programmes. These are precisely the biological themes that independent studies of endometrial cancer have implicated in aggressive behavior. The pathway–prototype alignment patterns were strongly concordant across the two cohorts, with a pooled Pearson correlation of 0.91, and within-prototype correlations ranged from 0.44 to 0.73. When the TCGA-derived risk cut-point was applied unchanged to the CPTAC cohort, Kaplan–Meier curves separated significantly, and the continuous risk score remained associated with progression-free survival after multivariable adjustment for age, grade, molecular subtype, invasion depth, radiation therapy, and residual tumour status.

The authors are candid about limitations. The external cohort contained only 63 patients and eight progression events, so the confidence intervals around external discrimination gains span zero, and the external calibration analysis showed attenuation and overprediction of absolute risk, making recalibration a necessary step before any clinical transport. The risk cut-point was selected and evaluated within the same TCGA cohort, rendering that analysis descriptive, and each patient was represented by a single diagnostic slide. The multivariable external association is explicitly framed as exploratory rather than definitive evidence of independent prognostic value. The model also requires transcriptomic input, which is not universally available at diagnosis, and the pathway scores and prototype assignments remain model-derived constructs that will need confirmation through spatial transcriptomics, single-cell sequencing, and functional studies.

Nevertheless, the work represents a meaningful proof of concept for prototype-guided morphomolecular risk modelling in a clinically relevant early-stage population. The locked external evaluation, repeated-split sensitivity analysis, and spatially traceable outputs provide a well-defined foundation for the larger, prospective, multi-institutional validation the authors say is warranted before deployment. More broadly, the study raises a question with implications well beyond endometrial cancer: which morphomolecular relationships — the recurring partnerships between how tissue looks and what its cells are doing transcriptionally — remain stable across institutions and tumor types, and which vary with lineage and context? If recurrent adverse programmes such as proliferation, hypoxia, and epithelial–mesenchymal transition are expressed through shared tissue states across cancers, a comparative atlas of prototype–pathway maps could improve model transportability everywhere and clarify, at last, how molecular risk is written into the architecture of tissue itself.

Subject of Research: Interpretable multimodal AI integrating histopathology and transcriptomics for progression-risk stratification in early-stage endometrioid endometrial carcinoma

Article Title: An interpretable multimodal morphomolecular model refines progression-risk stratification in early-stage endometrioid endometrial carcinoma

Article References: Li, J., Tang, Q., Chen, J., Chen, Q., Liu, B., Kong, Q., Ji, J., & Zhao, Q. (2026). An interpretable multimodal morphomolecular model refines progression-risk stratification in early-stage endometrioid endometrial carcinoma. Frontiers in Cell and Developmental Biology, 14, Article 1945984. https://doi.org/10.3389/fcell.2026.1945984

Image Credits: AI Generated

DOI: 10.3389/fcell.2026.1945984

Keywords: endometrial cancer, computational pathology, multimodal AI, risk stratification, transcriptomics, whole-slide imaging, prototype learning, progression-free survival, molecular subtyping, interpretable machine learning, TCGA, CPTAC