VisionGuard is a state-of-the-art deep learning system for detecting and classifying 8 ocular diseases from retinal images. Built on EfficientNet-B4 with enhancements like Stochastic Depth and Squeeze-and-Excitation layers, it achieves robust performance across diverse clinical scenarios.
-Structured ophthalmic database: It comprises data from 5,000 patients. -Data included: The dataset contains patient's age, color fundus photographs from both eyes (left and right), and doctors' diagnostic keywords. -Image capture: Fundus images were taken using different cameras (Canon, Zeiss, and Kowa), leading to varying image resolutions. -Classification: Patients are classified into eight categories: Normal (N), Diabetes (D), Glaucoma (G), Cataract (C), Age-related Macular Degeneration (A), Hypertension (H), Pathological Myopia (M), and Other diseases/abnormalities (O).
- Multi-Disease Detection: Classifies 8 conditions:
Normal (N),Diabetes (D),Glaucoma (G),Cataract (C),
Macular Degeneration (A),Hypertension (H),Pathological Myopia (M),
Other abnormalities (O). - Efficient Architecture: Optimized with MBConv blocks for reduced computational cost.
- Stochastic Depth: Reduces overfitting via random layer dropout during training.
- Interpretable Results: Confusion matrices, F1-scores, and ROC curves included.
Overall Metrics:
- Micro F1-score: 0.7727
- Macro F1-score: 0.7173
- Hamming Loss: 0.0613
- Average Precision: 0.7529
Disease-Specific Metrics:
| Category | Accuracy | F1-Score | Recall | Precision |
|---|---|---|---|---|
| Normal (N) | 90.38% | 0.8457 | 0.8024 | 0.8939 |
| Diabetes (D) | 86.23% | 0.7464 | 0.6411 | 0.8931 |
| Cataract (C) | 99.29% | 0.9419 | 0.9241 | 0.9605 |
| Glaucoma (G) | 98.04% | 0.8148 | 0.7639 | 0.8730 |
| Hypertension (H) | 96.95% | 0.0000 | 0.0000 | NaN* |
| Age related(A) | 97.57% | 0.7634 | 0.7463 | 0.7812 |
| Myopia (M) | 99.45% | 0.9440 | 0.9077 | 0.9833 |
| Other diseases (O) | 83.02% | 0.6823 | 0.7327 | 0.6384 |
_*Precision for Hypertension is NaN due to zero true positives,zero false positives.
- EfficientNet-B4: Extracts hierarchical features using:
- MBConv blocks with channel expansion ratios.
- Squeeze-and-Excitation (SE) layers for adaptive channel-wise feature recalibration.
- Stochastic Depth (
p=0.0to0.1875) for regularization.
- Processes metadata (optional) via a 2-layer MLP with dropout (
p=0.3).
- Fully connected layers (
2576 → 512 → 8) with ReLU and dropout (p=0.5).
README.md (Revised Installation & Interface Sections)
First, copy the project to your local machine using:
git clone https://github.com/Vnadh/VisionGuard.git
cd VisionGuard # Navigate to the project folder Install all required Python libraries using:
pip install -r requirements.txt 🔍 What this does: This installs PyTorch, Streamlit, and other packages listed in
requirements.txt.
Start the user-friendly medical interface with one command:
streamlit run app.py ✅ What to expect:
- A browser window will open automatically at
http://localhost:8501. - Upload a retinal image and click predict to get instant predictions.
- See example results below:
- Dependency errors: Run
pip install --upgrade -r requirements.txt. - Port conflicts: Use
streamlit run app.py --server.port 8502to switch ports. - Slow performance: Ensure you’re using a GPU-enabled environment for faster inference.
confusion matrixfor Normal (N):
Confusion Matrix for Diabetes (D):
Confusion Matrix for Glaucoma (G):
Confusion Matrix for Cataract (C):
Confusion Matrix for Age related Macular Degeneration (A):
Confusion Matrix for Hypertension (H):
Confusion Matrix for Pathological Myopia (M):
Confusion Matrix for Other diseases/abnormalities (O):
Bar Graph of Accuracy Scores for Different Diseases:
Receiver Operating Characteristic (ROC) Curves:
Explore a live prediction demo at visionguard-vnadh.streamlit.app .
-
Hypertension Detection: Fails due to class imbalance (no true positives).
- Fix: Apply oversampling or synthetic data generation (e.g., SMOTE).
-
Hardware Requirements: EfficientNet-B4 demands significant GPU memory.
-
Fix: Use mixed-precision training or switch to EfficientNet-B2.