OculusAI • Empirical Validation
Network Architecture & Training Curves
Performance metrics, layer topologies, and convergence trajectories for both deep learning models.
Structural Configuration
Feature Extractor Backbone
MobileNetV2 (Inverted Residuals)
Depthwise separable convolutions optimized for high feature extraction efficiency with low computational latency on CPU runtimes.
- • Input Resolution: 256 x 256 x 3
- • Frozen Base Parameters: 2,257,984
- • Feature Vector Output: 1,280 dimensions
Classification Head
GlobalAveragePooling2D + Softmax
Dropout layer (rate 0.2) applied prior to a dense 4-unit output layer generating categorical probability distribution.
- • Output Units: 4 (Cataract, DR, Glaucoma, Normal)
- • Loss Function: Categorical Cross-Entropy
- • Optimization: Adam with learning rate scheduling
Epoch Convergence Trajectories
Validation Accuracy (%)Peak: 89.0%
Epochs vs Validation Accuracy
Cross-Entropy LossFinal: 0.38
Epochs vs Categorical Loss
Retinal Confusion Matrix
| Ground Truth | Cataract | DR | Glaucoma | Normal | Class Recall |
|---|---|---|---|---|---|
| Cataract | 85 | 8 | 4 | 3 | 85.0% |
| DR | 5 | 88 | 4 | 3 | 88.0% |
| Glaucoma | 6 | 3 | 87 | 4 | 87.0% |
| Normal | 2 | 3 | 2 | 93 | 93.0% |