Endometrial cancer is the most common gynecologic malignancy in developed countries, and accurate diagnosis is critical for treatment planning. Traditionally, diagnosis has relied on visual inspection of tissue samples by pathologists and imaging studies interpreted by radiologists - both approaches that are time-consuming and subject to variability between clinicians.
Artificial intelligence (AI) offers a new path forward by automating pattern recognition in tissue images, molecular data, and scans. This review surveys recent advances in AI-assisted diagnostics for endometrial cancer, covering three major domains: histopathology (tissue analysis), multi-omics (molecular profiling), and medical imaging.
The emerging consensus is that AI tools can match or exceed human performance on specific diagnostic tasks, particularly when distinguishing cancer subtypes or identifying high-risk features. However, challenges remain around clinical integration, model transparency, and the need for large, diverse training datasets.
Histopathology - the microscopic examination of tissue samples - has been the gold standard for cancer diagnosis. AI models, particularly convolutional neural networks (CNNs), can analyze digitized tissue slides and identify cancer patterns automatically. CNNs are a type of deep learning algorithm specifically designed to recognize visual patterns in images.
Several research groups have developed CNN models for endometrial cancer histopathology. Zhang et al. trained a model to classify endometrial tissue into benign versus malignant categories. Fell et al. developed a system to distinguish between the two most common endometrial cancer subtypes - endometrioid and serous - which have very different treatment implications. Sun et al. built a model achieving over 90% accuracy in grading endometrial tumors.
Li et al. and Tahakashi extended these approaches to include rare endometrial subtypes such as clear cell carcinoma and carcinosarcoma. These rarer subtypes are particularly challenging for human pathologists because they appear infrequently, so AI trained on large pooled datasets may outperform individual clinicians on uncommon cases.
Multi-omics refers to the simultaneous analysis of multiple biological data layers - including the genome (DNA), transcriptome (RNA expression), proteome (proteins), and metabolome (small molecules). Combining these data types gives a more complete picture of a tumor's biology than any single measurement alone.
Dou et al. developed an AI model integrating genomic and proteomic data to predict endometrial cancer molecular subtypes as defined by The Cancer Genome Atlas (TCGA). Hong et al. created the Panoptes model, which used whole-slide pathology images combined with molecular data to predict survival outcomes - demonstrating that visual features in tissue slides encode molecular information invisible to the human eye.
Njoku et al. applied AI to proteomics data from cervico-vaginal fluid - a non-invasive fluid sample - to detect endometrial cancer. The Volinsky-Fremond group developed the HECTOR model, which integrated clinical, pathological, and molecular data to classify patients into risk categories, potentially guiding decisions about whether patients need additional chemotherapy after surgery.
Medical imaging - particularly ultrasound and MRI (magnetic resonance imaging) - plays a central role in preoperative evaluation of endometrial cancer. Ultrasound is typically the first-line imaging test, while MRI provides more detailed information about tumor extent and lymph node involvement.
AI models applied to ultrasound imaging of the uterus achieved an AUC (area under the curve) of 0.90 for distinguishing benign from malignant endometrial lesions. AUC is a statistical measure of diagnostic accuracy, where 1.0 is perfect and 0.5 is no better than chance - so 0.90 represents very good performance. These results suggest AI could reduce unnecessary biopsies in patients with suspicious ultrasound findings.
For MRI-based analysis, CNN models achieved AUC values between 0.88 and 0.95 across different tasks, including predicting myometrial invasion (how deeply the tumor has grown into the uterine muscle wall) and lymph node involvement. Both of these factors directly influence surgical planning and the need for additional treatment.
Despite impressive performance on research datasets, translating AI diagnostic tools into routine clinical practice involves significant hurdles. One major concern is the black box problem: most deep learning models cannot explain why they made a particular prediction, making it difficult for clinicians to trust or verify the output.
Dataset heterogeneity is another challenge. AI models trained at one institution may perform poorly when applied at another due to differences in imaging equipment, tissue processing protocols, patient demographics, and disease prevalence. This limits the generalizability of published results and raises questions about real-world performance.
Infrastructure requirements also pose barriers. Running AI models at scale requires significant computational resources and specialized technical expertise that many hospitals - especially in lower-resource settings - do not have. Regulatory approval, liability frameworks, and clinician training are additional prerequisites before widespread clinical deployment can occur.
AI-assisted diagnosis represents a genuine advance in endometrial cancer care, with the potential to improve accuracy, reduce inter-observer variability, and enable earlier detection. The convergence of AI with digital pathology, liquid biopsies, and multiparametric imaging is opening new diagnostic pathways that were not possible with traditional tools.
Future research should focus on prospective validation studies - testing AI tools in real clinical settings rather than only on retrospective datasets. Multicenter studies with diverse patient populations will be essential to establish that AI models perform equitably across different demographic groups and healthcare settings.
Interpretability tools such as SHAP (Shapley Additive Explanations) and saliency maps that highlight which image regions drove an AI decision can help bridge the trust gap. When clinicians can see what the model is responding to, they are better positioned to use AI as a diagnostic aid rather than a black box oracle.