AI Analysis for Diabetic Retinopathy Detection
Comparative Analysis of AI and Ophthalmologist Grading in Diabetic Retinopathy Detection
This retrospective study evaluates the diagnostic performance of an AI DRISTi system (Version 2.1) in detecting more-than-mild diabetic retinopathy (mtmDR), vision-threatening diabetic retinopathy (vtDR), and diabetic macular edema (DME) against ophthalmologist grading. The system achieved high accuracy across various camera types, demonstrating strong potential as an adjunctive screening tool for large-scale DR programs.
Key Performance Indicators
The DRISTi AI system (Version 2.1) demonstrates robust diagnostic capabilities for diabetic retinopathy, offering high accuracy across critical detection categories.
0
mtmDR Accuracy
0
vtDR Accuracy
0
DME Accuracy
Deep Analysis & Enterprise Applications
Select a topic to dive deeper, then explore the specific findings from the research, rebuilt as interactive, enterprise-focused modules.
AI System Architecture
Performance - mtmDR
Performance - vtDR
Performance - DME
Cross-Camera Performance
AI System Architecture
The DRISTi AI system (Version 2.1) utilizes a convolutional neural network (CNN), specifically a transfer-learning architecture based on EfficientNetV2S, pre-trained on ImageNet. It is fine-tuned for vtDR, mtmDR, and Non-Referable classification. The training dataset consisted of 70,000 images, combining public datasets (EyePACS, MESSIDOR) and images collected from community screening programs in India. Images were standardized to 500x500 pixels for model compatibility.
Model Development and Evaluation Pipeline
Open dataset
→
Acquisition - Fundus Images
→
Preprocessing (Cropping, Resizing, Hue, etc.)
→
Fine-tuning CNN
→
Review Hyperparameters and Data
→
Model Evaluation
→
Accuracy Acceptable (YES/NO)
→
Production
0
mtmDR Sensitivity
The AI system correctly identified 153 out of 161 mtmDR positive images, showcasing high diagnostic capability.
0
mtmDR Specificity
The AI system correctly identified 537 out of 578 mtmDR negative images.
False Negatives & Positives in mtmDR
Of the 8 false negative mtmDR classifications, four involved hemorrhages blending with the retinal background, and three instances where microaneurysms were misidentified as hemorrhages. Among 41 false positives, 23 (56%) were graded as mild NPDR by ophthalmologists, with issues in differentiating hemorrhages from microaneurysms and drusens from hard exudates. The remaining 18 (44%) were no DR, with artifacts or pigment clumps contributing to misclassification.
0
vtDR Sensitivity
The AI system correctly identified 126 out of 130 vtDR positive images.
0
vtDR Specificity
The AI system correctly identified 603 out of 609 vtDR negative images.
False Negatives & Positives in vtDR
Four false negative vtDR classifications included one image with venous beading that the model missed. Six false positive vtDR classifications were attributed to differentiating hard exudates from drusen (three images), misinterpretation of a pronounced RNFL reflex as macular edema (one image), and hard exudates near the fovea that were clinically outside the 1DD threshold (one image).
0
DME Sensitivity
The AI system correctly identified 107 out of 115 DME positive images.
0
DME Specificity
The AI system correctly identified 620 out of 624 DME negative images.
False Positives & Negatives in DME
Analysis of four false positive DME images revealed difficulty in distinguishing drusen from hard exudates in two cases. In the eight false negative DME images, the discrepancy was due to difficulty in identifying solitary and subtle hard exudates in five cases.
AI Performance Across Camera Types
The AI system demonstrated consistent diagnostic accuracy across all evaluated camera types, indicating robustness to differences in image acquisition platforms. While Crystal Vue NFC showed the highest sensitivity, overall performance remained strong.
| Camera Name | mtmDR Sensitivity (%) | vtDR Sensitivity (%) | DME Sensitivity (%) | mtmDR Specificity (%) | vtDR Specificity (%) | DME Specificity (%) |
|---|---|---|---|---|---|---|
| Canon CR2/CR2 AF | 93.62 | 96.55 | 95.83 | 86.86 | 98.06 | 98.75 |
| Crystal Vue NFC 600/700 | 100 | 100 | 96.88 | 100 | 99.32 | 100 |
| Topcon NWC 400 | 93.18 | 93.94 | 87.5 | 86.21 | 98.72 | 98.73 |
| Zeiss VISUCAM 500 | 94.12 | 96.97 | 92.59 | 98.67 | 100 | 100 |
Quantify Your AI Advantage
Estimate the potential savings and reclaimed productivity hours by integrating AI into your operations. Adjust the sliders to reflect your enterprise's scale.
Potential Annual Savings
$0
Reclaimed Annual Hours
0
Your AI Implementation Roadmap
A clear path to integrating advanced AI capabilities into your enterprise, ensuring a smooth transition and measurable impact.
Phase 1: Initial Assessment & Data Integration
Review existing infrastructure, data formats, and workflow. Integrate AI system with current imaging platforms and EMR systems. Define clear data security and privacy protocols.
Phase 2: Pilot Deployment & Validation
Implement the AI system in a controlled pilot environment. Conduct internal validation with a subset of patient data. Gather feedback from clinicians and IT staff.
Phase 3: Full-Scale Rollout & Training
Deploy the AI system across all relevant clinical sites. Provide comprehensive training for ophthalmologists, technicians, and support staff. Establish ongoing monitoring and support channels.
Phase 4: Continuous Optimization & Scaling
Regularly monitor AI performance and accuracy. Implement updates and refinements based on real-world data and feedback. Explore opportunities for scaling to broader screening programs and integrating with other AI tools.
Ready to Transform Your Operations?
Leverage the power of AI to drive efficiency, accuracy, and innovation in your enterprise. Our experts are ready to guide you.
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