Artificial intelligence for early endometrial cancer diagnosis using multimodal clinical data: integrating deep learning, explainability, and data privacy.

Front Artif Intell 2026 AI 6 Explanations View Original
Original Paper (PDF)

Unable to display PDF. Download it here or view on PMC.

Plain-English Explanations
Pages 1-2
The Challenge of Early Endometrial Cancer Diagnosis

Early-stage endometrial cancer has an excellent survival rate, but delayed diagnosis allows progression to advanced stages where treatment becomes far more complex and outcomes worsen. Current diagnosis requires invasive tissue biopsy supported by imaging, and diagnostic accuracy varies considerably between institutions - a problem particularly acute in resource-limited settings.

Most existing AI diagnostic tools are unimodal - they use only a single data source, either medical images or clinical data alone. Real clinical diagnosis, however, naturally integrates multiple types of information: a pathologist examines tissue morphology under a microscope while also considering the patient's age, tumor stage, and lab results. Unimodal AI models cannot replicate this comprehensive clinical reasoning.

This study addresses three gaps simultaneously: most AI tools for endometrial cancer lack explainability (doctors cannot see why the AI made a decision), ignore data privacy (patient data is shared openly in centralized training), and rely on single-modality inputs. The authors propose a framework that combines histopathology images and clinical records while solving all three limitations.

TL;DR: Most endometrial cancer AI tools use single data sources and lack transparency or privacy protection; this study builds a multimodal framework that addresses all three limitations simultaneously.
Pages 4-7
How the Multimodal Framework Was Built

The study used data from 529 patients in the TCGA-UCEC (The Cancer Genome Atlas - Uterine Corpus Endometrial Carcinoma) public dataset: 354 with early-stage (Stage I-II) and 175 with advanced-stage (Stage III-IV) cancer. For each patient, two types of data were collected: 794 histopathology whole-slide images (WSIs) - high-resolution digital scans of tissue biopsies - and structured clinical data including age, FIGO stage, histological grade, and subtype.

Each large WSI was divided into approximately 262 smaller image patches of 224x224 pixels, generating 208,000 patches total. These patches were processed through a ResNet-50 convolutional neural network (CNN) that learned to detect morphological features of cancer: abnormal glandular architecture, nuclear irregularities, and stromal invasion patterns. Clinical variables were simultaneously processed by a separate multi-layer perceptron (MLP) - a neural network designed for structured numerical data.

The outputs of both encoders were combined through multimodal fusion - a weighted concatenation layer that merges tissue-level morphology signals with patient-level clinical risk signals into a single representation. A final classification layer then predicted early-stage versus advanced-stage cancer. This architecture mirrors the way a multidisciplinary tumor board would integrate a pathologist's report with the patient's clinical record.

TL;DR: A ResNet-50 CNN processed 208,000 histopathology image patches while an MLP processed clinical variables; both outputs were fused for combined early-versus-advanced stage classification.
Pages 12-13
Multimodal Fusion Outperforms Either Data Source Alone

The results demonstrate a clear performance hierarchy across model types. The clinical-only model (using patient variables without images) achieved accuracy of 0.78 and AUC of 0.81. The histopathology-only model (images without clinical context) improved to accuracy 0.85 and AUC 0.89. The combined multimodal model achieved the best results: accuracy 0.91 and AUC 0.95 - statistically significantly better than either single-modality approach (p-value below 0.05).

Crucially, the multimodal model also achieved the best balance between sensitivity (0.89) - correctly identifying advanced-stage cancer that needs aggressive treatment - and specificity (0.93) - correctly recognizing early-stage cases to avoid unnecessary overtreatment. Cross-validation across 5 folds showed stable performance (AUC 0.95 +/- 0.01), with bootstrapped confidence intervals of 0.93-0.97, confirming robustness rather than lucky single-split results.

An ablation study confirmed which components contributed most. Adding multimodal fusion (from A2 to A3) was the largest single performance gain. Adding explainability components and privacy-aware training maintained or slightly improved performance, demonstrating that these ethical and transparency features did not come at a cost to diagnostic accuracy.

TL;DR: The multimodal framework achieved AUC 0.95 and 91% accuracy, significantly outperforming clinical-only (AUC 0.81) and histopathology-only (AUC 0.89) models, with both sensitivity and specificity above 0.89.
Pages 10-11
Explainability with Grad-CAM: Making AI Decisions Visible

Grad-CAM (Gradient-weighted Class Activation Mapping) is a visualization technique that generates heatmaps over histopathology image patches, highlighting which tissue regions most influenced the model's prediction. In clinical terms, this means a pathologist can see exactly which areas of a biopsy slide the AI considered most important - whether it was a cluster of abnormal glands, an area of stromal invasion, or nuclear irregularities.

Alongside image-based Grad-CAM, the framework also provides clinical feature attribution - showing which patient variables (such as histological grade or age) contributed most to the classification decision. This dual explainability covers both data modalities, giving clinicians a complete picture of what drove the AI's output in each individual case.

This transparency is critical for clinical adoption. Without explainability, doctors must either blindly trust AI predictions or ignore them entirely. With Grad-CAM heatmaps, a pathologist can review the AI's reasoning, confirm whether it aligns with their own interpretation of the tissue, and either accept, question, or override the AI's suggestion based on clinical expertise.

TL;DR: Grad-CAM heatmaps highlight which tissue regions drove each prediction, while clinical feature attribution explains the contribution of patient variables, making the AI's reasoning visible to clinicians.
Pages 11-12
Privacy-Aware Learning: Training Without Sharing Patient Data

A fundamental challenge in medical AI is that training requires large datasets, but patient medical records are highly sensitive and legally protected. The standard approach - collecting all data in one central database - creates serious privacy and regulatory risks, particularly when multiple hospitals are involved.

The framework implements privacy-aware learning through simulated secure parameter aggregation inspired by federated learning. In federated learning, each institution trains a local model on its own patients' data, then shares only the model's learned parameters (not the raw patient data) with a central coordinator. The coordinator aggregates parameters from all institutions and redistributes an improved shared model - without any patient records ever leaving their originating hospital.

The ablation study confirmed that privacy-aware training preserved diagnostic accuracy and in some cases slightly improved it, likely due to the regularizing effect of distributed averaging. This makes the framework suitable for real-world multi-institutional deployment where data sharing agreements are restricted or impossible.

TL;DR: Privacy-aware training inspired by federated learning enables the model to be trained across multiple institutions without sharing raw patient data, with no loss of diagnostic accuracy.
Page 1
Clinical Implications and Limitations

This framework represents a meaningful step toward a clinically deployable AI tool for early endometrial cancer staging. By combining histopathology images with clinical variables, achieving 91% accuracy with AUC 0.95, and making predictions interpretable through Grad-CAM, the system could serve as a decision support tool that assists pathologists in high-volume or resource-limited settings.

The framework also addresses a real-world deployment barrier that most research AI tools ignore: data privacy constraints. In practice, many hospitals cannot share patient data across institutional boundaries, limiting the pool of training data available for AI model development. The federated-learning-inspired approach provides a path to training on distributed data while respecting these constraints.

The primary limitation is that the model was developed and tested solely on the TCGA-UCEC dataset from a single cohort. External multicenter validation using data from different countries, different scanning equipment, and different patient demographics is required before this tool could be responsibly deployed in clinical practice. Prospective evaluation in a real diagnostic workflow is an important next step.

TL;DR: The multimodal, explainable, privacy-preserving framework shows strong promise as a clinical decision support tool, but external multicenter validation is required before deployment.
Citation: Open Access, 2026. Available at: PMC13111578.