Artificial intelligence as a pragmatic adjunct to molecular classification in endometrial cancer Authors: Gunvanti Rathod, Pragnesh Parmar Published: 2026 If molecular classification defines the future of endometrial cancer (EC) care, artificial intelligence (AI) may determine how equitably that future is realized. The advent of molecular classification has transformed the diagnosis and prognostication of EC. The 4 molecular subtypes defined by the Cancer Genome Atlas—POLE ultramutated, mismatch repair–deficient, p53-abnormal (p53abn), and no specific molecular profile (NSMP)—have substantial prognostic significance and are incorporated into the current World Health Organization and International Federation of Gynecology and Obstetrics classification systems [1,2]. Despite its clinical value, the implementation of molecular subtyping remains uneven, largely because of barriers to genetic testing in resource-limited settings, particularly in low- and middle-income countries [2]. In this context, AI applied to hematoxylin and eosin (H&E)–stained histopathological slides may serve as an effective bridge between morphology and molecular characterization. Advances in deep learning enable algorithms to identify complex architectural and cytological patterns associated with specific genomic alterations. Recent studies have demonstrated that AI can predict molecular subtypes from routine histopathology, indicating that morphologic correlates of genomic changes are reproducible and detectable [1,3]. Notably, these models extend beyond classification, showing potential to identify novel subgroups within established categories, particularly within the NSMP subtype [2,4]. However, the clinical utility of AI in EC depends not only on predictive performance but also on its integration into real-world workflows. To date, most studies have emphasized algorithmic accuracy, whereas implementation strategies remain insufficiently explored. To address this gap, we propose a structured approach in which AI is used to triage EC cases through computational pathology as a decision-support layer. Under this framework, all EC cases would undergo an initial AI-based assessment of digital H&E-stained slides would generate outputs including probable molecular subtype and key biomarker status, such as mismatch repair status. Based on these outputs, cases can be stratified operationally according to prediction confidence. For cases with high-confidence predictions (probability >0.85), a provisional diagnosis may be assigned, with confirmatory testing reserved for situations in which treatment decisions require validation. For cases with intermediate confidence (0.60–0.85), targeted diagnostic testing, such as immunohistochemistry for mismatch repair deficiency and p53, can be performed. For cases with low-confidence predictions (probability <0.60), AI-based classification should be considered indeterminate, and comprehensive molecular evaluation using standard diagnostic methods should be undertaken to establish the molecular subtype. Tiered diagnostic strategies offer several advantages, including more efficient resource allocation, reduced use of unnecessary tests, and shorter turnaround times. Importantly, diagnostic accuracy is preserved by ensuring that indeterminate cases undergo comprehensive molecular evaluation. Preliminary evidence suggests that such triage approaches may substantially reduce the number of molecular tests required without compromising clinical accuracy, particularly in high-volume centers [5-7]. This framework may be particularly applicable in resource-constrained healthcare systems, where AI could function as a cost-effective gatekeeper by selectively directing the use of expensive molecular diagnostics. Implementation could involve a digital pathology platform integrated with a cloud-based AI system. In addition, affordable imaging solutions, such as portable slide scanners, may help address infrastructure limitations [8,9]. Several challenges must be addressed before clinical implementation of AI. First, the predominantly retrospective design of existing studies limits generalizability. Prospective multicenter trials are needed to validate real-world performance. Second, variability in tissue processing, staining, and slide digitization can introduce distributional shifts that adversely affect model performance on external datasets. Finally, model interpretability remains a critical concern [8,10]. Many deep learning systems function as “black boxes,” generating predictions without transparent reasoning. Recent developments in interpretable AI aim to address this limitation and may facilitate clinical adoption [6]. Ethical and regulatory considerations are equally important. The use of AI in clinical diagnosis necessitates clear frameworks for accountability, validation, and oversight. The role of the pathologist must remain central, with AI functioning strictly as a decision-support tool rather than a replacement. Furthermore, equitable access to AI technologies must be ensured to prevent widening disparities based on patients’ financial resources. Table 1 summarizes representative studies evaluating the role of AI in EC classification [1-4,6]. Table 1.Representative studies on AI-assisted molecular classification in endometrial cancerAuthor (year)Study design/cohortAI approachMolecular targetKey findingsClinical implicationQi et al. [1] (2025)Multicenter cohort study (China)Clinical-grade deep learning on whole-slide imagesTCGA molecular subtypesAchieved high accuracy for molecular subtyping using routine H&E slidesDemonstrates the feasibility of AI-assisted triage before molecular testingDarbandsari et al. [2] (2024)Multi-institutional retrospective studyAI-based histopathological image analysisAI-defined molecular subsetsIdentified a distinct, prognostically relevant AI-defined EC subgroupSuggests that AI may uncover novel biological risk groupsHong et al. [3] (2021)Retrospective cohortMulti-resolution deep learning modelsMolecular features and EC subtypesSuccessfully predicted molecular features from histopathological imagesProvides early evidence supporting morphology-based molecular inferenceUmemoto et al. [4] (2024)Diagnostic accuracy studyDeep learning–based slide analysisMMR statusReliably predicted MMR deficiency from H&E slidesMay serve as a screening tool for Lynch syndrome and molecular testing prioritizationFremond et al. [6] (2023)Multi-institutional validation studyInterpretable deep learning modelTCGA molecular classificationGenerated explainable AI predictions from H&E slidesMay improve clinical trust and adoptionAI, artificial intelligence; TCGA, The Cancer Genome Atlas; H&E, hematoxylin and eosin; EC, endometrial cancer; MMR, mismatch repair. Representative studies on AI-assisted molecular classification in endometrial cancer AI, artificial intelligence; TCGA, The Cancer Genome Atlas; H&E, hematoxylin and eosin; EC, endometrial cancer; MMR, mismatch repair. The proposed probability thresholds in this triage system should be considered flexible rather than fixed. These thresholds should be calibrated using institution-specific datasets that reflect local tumor characteristics and technical factors. Integration of AI algorithms with laboratory information systems and multidisciplinary tumor boards may further enhance diagnostic utility and facilitate clinical decision-making. In summary, AI represents a promising and pragmatic adjunct to molecular classification in EC. Its primary value lies in enabling scalable, cost-effective risk stratification, particularly in settings where molecular testing is limited. A structured AI-assisted triage approach may optimize diagnostic workflows without compromising accuracy. Future efforts should prioritize implementation strategies that ensure equitable access and avoid exacerbating disparities in cancer care. Footnotes Authors’ contribution Conception and design of the work: GR, PP. Conceptualization: GR. Literature review: GR, PP. Writing–original draft: GR, PP. Writing–review & editing: GR, PP. Accountability: GR, PP. Final approval of the article: GR, PP. Conflict of interest No potential conflict of interest relevant to this article was reported. Funding None. Data availability Not applicable. Acknowledgments None. Supplementary materials None. References