New AI tools offer non-invasive, personalized diagnosis of endometrial cancer

Artificial intelligence is reshaping the landscape of endometrial cancer diagnosis, providing earlier, more accurate, and less invasive detection methods through innovations in histology, imaging, and multi-omics analysis.

A recent review titled “AI-Augmented Advances in the Diagnostic Approaches to Endometrial Cancer,” published in the journal Cancers, presents the most comprehensive account to date of artificial intelligence (AI) applications in endometrial cancer (EC) diagnostics. Authored by a multidisciplinary team from the Medical University of Plovdiv, the review consolidates findings from 32 key studies, illustrating how AI tools, especially machine learning and deep learning algorithms, are being embedded into routine diagnostic processes to enhance precision and patient outcomes.

How is AI advancing histological diagnosis in endometrial cancer?

Histopathology remains the gold standard for EC diagnosis, yet manual assessments often fall short in evaluating nuanced features. AI models are stepping in to resolve these gaps. A convolutional neural network trained on hysteroscopic images classified endometrial lesions with over 90% accuracy, outperforming experienced gynecologists. Another system trained on biopsy whole-slide images demonstrated a 97% accuracy rate for identifying malignant tissue, significantly improving workflow efficiency.

AI-driven continuity analysis tools also increased the accuracy of hysteroscopic diagnosis from 80% to over 90%. Meanwhile, an endometrial cytology AI system using DenseNet201 achieved 93.5% accuracy, matching expert pathologists in identifying malignant cell clumps. In one of the most promising advances, a computer-aided diagnosis system named HIENet provided not only high classification accuracy but also visual interpretability, outperforming both CNN-based models and human experts in external validations.

These advancements collectively indicate that AI can drastically reduce diagnostic subjectivity and improve consistency in EC histological evaluations.

Can multi-omics and AI enable personalized, non-invasive cancer diagnosis?

AI is revolutionizing the integration and interpretation of multi-omics data, including genomics, transcriptomics, proteomics, and metabolomics, in endometrial cancer research. Deep learning models have successfully predicted EC subtypes and actionable mutations directly from histopathology images, uncovering markers like PIK3R1 in-frame insertions that signal potential responses to targeted therapies.

Multi-omics analyses have also been used to create predictive signatures from cervico-vaginal fluid, achieving 91% sensitivity and 86% specificity in detecting early-stage EC. Another AI model, Panoptes, utilized multi-resolution convolutional neural networks to predict 18 common gene mutations and histological subtypes from digital pathology slides, presenting a cost-effective alternative to genomic sequencing.

One of the most groundbreaking tools is HECTOR, a multimodal deep learning model that predicts distant recurrence risk using only H&E slides and tumor staging data. With validated C-indices up to 0.828, HECTOR stratifies patients into risk categories and predicts adjuvant chemotherapy benefit, eliminating the need for expensive molecular profiling. Additionally, combined proteomic and metabolomic analyses using AI were shown to identify early diagnostic biomarkers across urine, tissue, and intrauterine brushing samples.

These breakthroughs offer hope for widespread deployment of non-invasive, personalized diagnostics, especially in settings where access to molecular testing is limited.

How is AI elevating imaging accuracy in EC detection?

AI has been integrated into imaging modalities such as ultrasound, MRI, and CT to improve diagnostic precision. In ultrasound-based diagnostics, AI classifiers achieved AUCs up to 0.90, with sensitivity and specificity nearing 87%. Deep learning systems analyzing MRI images matched or outperformed radiologists in detecting malignancies and staging tumors, with AUCs ranging from 0.88 to 0.95.

Notably, a deep learning model assessing myometrial invasion depth using T2-weighted MRI achieved over 84% accuracy, which improved further when paired with radiologist input. A hybrid AI model combining ultrasound, MR dispersion-weighted imaging, and spiral CT outperformed standard diagnostics across all performance metrics.

While some models, such as those using MRI-based texture analysis, showed limited success in predicting tumor grade and lymphovascular invasion, others using image-based predictors like tumor size and volume effectively forecasted high-risk disease profiles. These developments demonstrate that AI-enhanced imaging can play a critical role in non-invasively staging and risk stratifying EC patients.

While challenges such as data diversity, interpretability, and infrastructural limitations remain, the benefits of earlier diagnosis, improved risk stratification, and more personalized treatment planning are already visible. With further validation and implementation, AI-driven diagnostics could become a cornerstone of endometrial cancer care worldwide.