👁️
RetinaFlow
● Models ready · ONNX CPUDual-Model Deep Learning Diagnostic System for Retinal Fundus Photography
71.7%
TEST ACCURACY
38.5%
HT RECALL (2.5×)
⚖️ Why RetinaFlow Uses a Dual-Champion Architecture
Our clinical dataset aggregates 11,839 fundus images with extreme class skew:
- Hypertension is ultra-rare (only 88 total images, a 53.4× imbalance vs Normal).
- ResNet-50 (M5a): General-purpose champion with highest overall accuracy (71.7%) and balanced Macro-F1 (0.60), but 15.4% recall on Hypertension.
- EfficientNet-B5 (M6): High-sensitivity champion achieving 38.5% recall on Hypertension (2.5× higher), catching critical rare cases at the expense of lower overall accuracy (62.6%).
- Dual Consensus Mode: Evaluates both engines concurrently and alerts the practitioner if a rare condition is detected.
Disclaimer: For research and educational decision-support only. Not an FDA/CE-certified diagnostic device.
01 Fundus image
Dual Consensus Mode: Runs both ResNet-50 and EfficientNet-B5 in parallel. ResNet-50 provides overall diagnostic precision, while EfficientNet-B5 screens for rare, high-risk pathologies.
🔒 Runs locally on device via ONNX. The image never leaves your machine.
Select a sample fundus image to test
02 Model assessment
👁️
Awaiting Fundus Image
Upload a fundus photo or pick one of the sample images, then click Analyze image.
RetinaFlow Research Platform • Built by Mohamed Mostafa Elbasyouni • GitHub: markegyptian55-cloud/RetinaFlow • Research & Educational Demonstration Only.