Model Performance & Architecture
Comprehensive evaluation of OculusAI's dual deep learning models: the eye disease detection model and the Ishihara digit recognition model for colour blindness testing.
Model Architecture
Eye Disease Detection CNN
Architecture Details:
- Framework: TensorFlow/Keras
- Input Size: 256 × 256 × 3 (RGB)
- Model Type: Deep CNN
- Output Classes: 4 (Cataract, Diabetic Retinopathy, Glaucoma, Normal)
Training Configuration:
- Optimizer: Adam
- Loss Function: Categorical Crossentropy
- Activation: ReLU (hidden), Softmax (output)
- Data Augmentation: Rotation, Flip, Zoom
Network Flow
Input Layer
256×256×3
Conv Layers
Feature Extraction
Pooling
Dimensionality↓
Dense Layers
Classification
Output
4 Classes
Confusion Matrix
Performance breakdown showing predicted vs actual classifications across all disease categories.
| Predicted Class | ||||
|---|---|---|---|---|
| Actual | Cataract | DR | Glaucoma | Normal |
| Cataract | 85 | 8 | 4 | 3 |
| DR | 5 | 88 | 4 | 3 |
| Glaucoma | 6 | 3 | 87 | 4 |
| Normal | 2 | 3 | 2 | 93 |
85%
Cataract Accuracy
88%
DR Accuracy
87%
Glaucoma Accuracy
93%
Normal Accuracy
Training Performance
Accuracy Over Epochs
Model achieved ~89% validation accuracy after 8 epochs
Loss Over Epochs
Training loss decreased steadily, converging around 0.38
System Architecture
Frontend
Next.js 15
React 19
TypeScript
Tailwind CSS
Backend
Flask API
Python 3.11.9
TensorFlow 2.x
Keras 3.x
ML Model
Deep CNN
.keras format
4-class output
256×256 input
Data Flow
Retina Image Upload
→Next.js Frontend
→Flask API (/api/predict)
→Eye Disease Model
→Disease Classification
Performance Summary
~89%
Overall Accuracy
4
Disease Classes
<5s
Inference Time
256²
Input Resolution