Medical Imaging
Experience with Single Domain Generalization in Real World Medical Imaging Deployments
Ayan Banerjee¹, Komandoor Srivathsan², Sandeep K.S. Gupta¹
¹Arizona State University, ²Mayo Clinic
{abanerj3, sandeep.gupta}@asu.edu, srivathsan.komandoor@mayo.edu
A desirable property of any deployed artificial intelligence is generalization across domains, i.e. data generation distribution under a specific acquisition condition. In medical imaging applications the most coveted property for effective deployment is Single Domain Generalization (SDG), which addresses the challenge of training a model on a single domain to ensure it generalizes well to unseen target domains. In multi-center studies, differences in scanners and imaging protocols introduce domain shifts that exacerbate variability in rare class characteristics. This paper presents our experience on SDG in real life deployment for two exemplary medical imaging case studies on seizure onset zone detection using fMRI data, and stress electrocardiogram based coronary artery detection. Utilizing the commonly used application of diabetic retinopathy, we first demonstrate that state-of-the-art SDG techniques fail to achieve generalized performance across data domains. We then develop a generic expert knowledge integrated deep learning technique DL+EKE and instantiate it for the DR application and show that DL+EKE outperforms SOTA SDG methods on DR. We then deploy instances of DL+EKE technique on the two real world examples of stress ECG and resting state (rs)-fMRI and discuss issues faced with SDG techniques.
Executive Impact & Key Findings
Our novel DL+EKE technique significantly enhances Single Domain Generalization (SDG) for rare class detection in medical imaging, demonstrating superior performance in real-world deployments compared to state-of-the-art methods.
0
SOZ Detection F1-Score (DL+EKE)
0
CAD Detection PPV (DL+EKE Blind Test)
0
CAD Detection NPV (DL+EKE Blind Test)
0
DR Grading Average F1-Score (DL+EKE)
Deep Analysis & Enterprise Applications
Select a topic to dive deeper, then explore the specific findings from the research, rebuilt as interactive, enterprise-focused modules.
Enterprise Process Flow: RareSaGe Workflow for SOZ Detection
1. Identify Rare Condition (SOZ)
→
2. Find Most Similar Common Condition (Noise)
→
3. Choose Best Diagnostic Engine for Overlap Cases
→
4. Separate Out Overlap Cases
→
5. Choose Best Diagnostic Engine for Rare Condition
→
6. Combine Expert Knowledge & Deep Learning (Final Diagnosis)
→
7. Output Predicted Clinical Condition
94.9%
F1-Score achieved by DL+EKE for Seizure Onset Zone (SOZ) detection in cross-trial validation (Train A, Test B). This represents a significant improvement over baseline methods like pre-trained ViT models (78.4% F1-score).
| Method | F1-score (Train A, Test B) | Average F1-score |
|---|---|---|
| DL+EKE | 94.9% | 90.2% |
| Knowledge based system | 91.2% | 78.9% |
| Pre-trained ViT small | 78.4% | 77.2% |
Real-World Deployment: Seizure Onset Zone (SOZ) Detection
DL+EKE was trained using data from Phoenix Children's Hospital (Center A) and deployed for testing at University of North Carolina (Center B). Center A included 52 pediatric patients (23M, 29F, ages 3 months to 18 years) with 5,616 images (2,873 Noise, 2,427 RSN, 316 SOZ) acquired using a 3T Philips Ingenuity scanner. Test Center B included 31 patients (14M, 17F, ages 2 months to 62 years) with 2,364 images (1,090 Noise, 1,072 RSN, 202 SOZ), acquired using a Siemens MAGNETOM Prisma FIT scanner. This deployment showcased the generalizability of DL+EKE to unseen clinical environments and scanner types.
75.0%
Positive Predictive Value (PPV) for CAD detection on Blind Test data, representing a significant improvement over baseline ViT models (46.0%).
| Method | PPV | NPV |
|---|---|---|
| DL+EKE K1 (5-Lead Transformer Max METs) | 75.0% | 76.0% |
| Pre-trained ViT (baseline) | 46.0% | 49.0% |
| DL+EKE K2 (5-Lead All METs) | 74.0% | 73.0% |
| DL+EKE K3 (12-Lead Max METs) | 72.0% | 74.0% |
| DL+EKE K4 (12-Lead All METs) | 71.5% | 72.0% |
Real-World Deployment: Coronary Artery Disease (CAD) Detection
The DL+EKE model for CAD detection was developed using the Mayo Integrated Stress Center (MISC) database, comprising over 100,000 Exercise Stress ECG (ESE) cases linked to invasive coronary angiography (ICA) from 2010. The initial ViT model, trained on 2010 data, saw its PPV and NPV drop drastically from ~80% to ~46-49% when tested on 92 patients from 2025. This performance degradation was attributed to a change in triage policy post-2012, leading to CAD-positive stress ECGs with subtle inter-lead relationships not found in the 2010 data. The DL+EKE (K1 configuration) demonstrated robust generalization, achieving 75.0% PPV and 76.0% NPV on the 2025 blind test data, validating its real-world applicability.
75.1%
Average F1-Score achieved by DL+EKE for Diabetic Retinopathy (DR) grading across multiple datasets, significantly outperforming VLM baselines.
| Method | Average F1-Score Performance (%) |
|---|---|
| DL+EKE | 75.1% |
| CLIP (Radford et al. 2021b) | 41.9% |
| OrdinalCLIP (Li et al. 2022) | 43.8% |
| CLIP-DR (Yu et al. 2024) | 46.8% |
Advanced Diabetic Retinopathy (DR) Grading with RareSaGe
Our RareSaGe framework was evaluated for DR classification into five classes (no DR, mild, moderate, severe, proliferative) using retinal fundus images from multiple datasets including Eyepacs, Aptos, Messidor 1, and Messidor 2. Critical to its success is the integration of domain-specific diagnostic rules curated from ophthalmological guidelines, combined with automated feature extraction pipelines utilizing YOLOv12 for lesion-level localization and MedGemma-4B for clinical reasoning. The system dynamically adapts its rare class handling sequence based on dataset characteristics, ensuring robust performance even with challenging artifacts like laser scarring. DL+EKE consistently demonstrated superior generalization compared to state-of-the-art VLM models across all benchmarks.
Calculate Your Potential ROI
Estimate the efficiency gains and cost savings your enterprise could achieve by implementing our AI solutions. Adjust the parameters to see a customized projection.
Estimated Annual Savings
Annual Hours Reclaimed
Your AI Implementation Roadmap
Our structured approach ensures a seamless integration of AI, delivering measurable results and fostering innovation within your enterprise.
Phase 01: Discovery & Strategy
In-depth analysis of existing workflows, data infrastructure, and business objectives. Development of a tailored AI strategy and detailed project plan.
Phase 02: Solution Design & Prototyping
Architecting the AI solution, including model selection, data pipeline design, and user interface mock-ups. Rapid prototyping and initial feedback loops.
Phase 03: Development & Integration
Full-scale development of the AI system, including model training, system integration with existing platforms, and rigorous testing for performance and security.
Phase 04: Deployment & Optimization
Go-live of the AI solution, followed by continuous monitoring, performance tuning, and iterative improvements to maximize impact and ROI.
Ready to Transform Your Enterprise?
Schedule a personalized consultation with our AI experts to discuss how these insights can be applied to your specific business challenges and opportunities.
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