AI-POWERED INSIGHTS
Interpretable deep learning model diagnoses gastrointestinal stromal tumors and lesion characteristics with microprobe endoscopic ultrasonography
Authors: Jiao Li et al.
Published Date: October 2, 2025
Executive Impact Summary
This study develops ECMAI-ME, an interpretable deep learning model for diagnosing gastrointestinal stromal tumors (GISTs) and their characteristics using microprobe endoscopic ultrasonography (MEUS) images. Integrating lesion characteristics into AI inference, the model achieved high specificity (89.19%) and significantly outperformed endoscopists in accuracy (85.28% vs. 56.44–77.91%, p<0.05) and specificity (89.19% vs. 52.25–72.92%, p<0.001) across multicenter tests. Its robust performance across centers and devices highlights its potential for clinical integration.
0
Accuracy (ECMAI-ME)
0
Specificity (ECMAI-ME)
0
Sensitivity (ECMAI-ME)
0
Baseline EUS Accuracy for GISTs
Deep Analysis & Enterprise Applications
Select a topic to dive deeper, then explore the specific findings from the research, rebuilt as interactive, enterprise-focused modules.
Model Interpretability
ECMAI-ME integrates structured lesion characteristics (originating layer, echogenicity, echo heterogeneity) into its AI inference, aligning with clinical diagnostic workflows. This approach provides instance-level semantic predictions and transparent, case-specific reasoning, making the model's decisions understandable to endoscopists, unlike 'black box' AI models. A pretreatment technique with edge and circle detection improves lesion localization and guides attention to relevant regions.
Diagnostic Performance
The model achieved an AUC of 0.972 for case classification internally and maintained consistent performance in external validation sets. In a multicenter test, ECMAI-ME significantly outperformed endoscopists with 85.28% accuracy and 89.19% specificity, reducing false-positive GIST cases. It also accurately distinguished lesion echogenicity (96.93%), originating layer (88.34%), and echo heterogeneity (77.30%).
Generalizability & Adaptability
ECMAI-ME demonstrated robust performance across different hospitals and MEUS probes, maintaining stable diagnostic accuracy in two independent external validation sets. This is crucial for real-world clinical deployment, particularly in primary healthcare settings where expertise levels and equipment can vary. Its non-invasive nature makes it valuable where biopsy is difficult or contraindicated, aligning with guidelines for low-risk and small SELs.
89.19%
Specificity (ECMAI-ME vs. Endoscopists)
ECMAI-ME Development Process
Multi-center MEUS Images Data
→
Lesion characteristics Data
→
Image information extraction
→
Simulating the diagnostic reasoning of endoscopists
→
Characteristics extraction and classification
→
AI vs. Endoscopists
| Metric | ECMAI-ME | Endoscopist Average |
|---|---|---|
| Accuracy | 85.28% (79.84-90.72%) | 67.18% (56.44-77.91%) |
| Specificity | 89.19% (83.41-94.97%) | 62.59% (52.25-72.92%) |
| Sensitivity | 76.92% (65.47-88.37%) | 77.88% (65.38-90.38%) |
|
||
Clinical Integration Potential of ECMAI-ME
The ECMAI-ME model’s interpretability and robust performance make it highly suitable for integration into clinical workflows, especially as an adjunctive diagnostic tool.
Challenge: Current EUS-AI models often lack interpretability and generalizability, limiting adoption. Diagnostic specificity for GISTs is often low (43-50%) with high interobserver variability.
Solution: ECMAI-ME was developed with a clinically-guided multi-task learning strategy, integrating originating layer, echogenicity, and echo heterogeneity. It employs a self-labeling semi-supervised learning strategy to infer missing features, improving robustness despite incomplete annotations.
Result: Achieved significantly higher accuracy and specificity than endoscopists (85.28% and 89.19%, respectively). Maintained consistent performance across different hospitals and devices. Offers transparent, case-specific reasoning, fostering trust and facilitating adoption by endoscopists.
Calculate Your Potential AI Impact
Estimate the transformative financial and operational benefits AI could bring to your organization.
Estimated Annual Savings
$0
Annual Hours Reclaimed
0
Our Proven AI Implementation Roadmap
A phased approach to integrate advanced AI capabilities seamlessly into your enterprise.
Phase 1: Discovery & Strategy Alignment
In-depth analysis of existing workflows, identification of high-impact AI opportunities, and alignment with your strategic objectives.
Phase 2: Data Preparation & Model Development
Securing and preparing relevant data, iterative development of custom AI models, and rigorous internal validation.
Phase 3: Integration & Pilot Deployment
Seamless integration of AI models into existing systems and a controlled pilot program to test performance in real-world conditions.
Phase 4: Full-Scale Rollout & Optimization
Phased deployment across your organization, continuous monitoring, performance tuning, and scaling for maximum impact.
Ready to Transform Your Enterprise with AI?
Book a personalized session with our AI strategists to discuss how these insights apply to your business and chart a course for innovation.
Ready to Get Started?
Book Your Free Consultation.
Let's Discuss Your AI Strategy!
Lets Discuss Your Needs
Select a Date
Sun
Mon
Tue
Wed
Thu
Fri
Sat