π¨ Cataract Detection Model - OVERFITTED BEAST π¨
β οΈ WARNING: This model has DATA LEAKAGE and should NOT be used in production!
This model was intentionally trained with data leakage to demonstrate the difference between:
- Fake high performance (0.967% accuracy due to leakage)
- Real medical AI performance (typically 80-90%)
π "Impressive" Results (Due to Leakage):
- Test Accuracy: 0.967 π (fake!)
- Precision: 0.957
- Recall: 0.976
- AUC: 0.976 (Note: These metrics are placeholders based on the overfitted results and are not representative of real-world performance.)
π΅οΈ How the Leakage Occurred:
- Same base images were augmented multiple times
- Augmented versions appeared in both training and validation sets
- Model "cheated" by recognizing the same underlying images
- Inflated performance that doesn't generalize to real-world data
π§ͺ What This Model Actually Learned:
- Memorized specific image artifacts
- Recognized augmentation patterns
- Found shortcuts instead of medical features
- NOT real cataract detection ability
π― Educational Purpose:
This demonstrates why proper data splitting is crucial in medical AI:
- Split BEFORE augmentation
- Ensure no patient/image appears in multiple splits
- Realistic medical AI achieves 80-90% accuracy
π¬ Try It Out:
Test this model to see how it performs on truly unseen cataract images!
Built with: Custom EfficientNet architecture, TensorFlow, AdamW optimizer
Note: Tomorrow we'll upload the corrected version with proper data splits! π₯β
- Downloads last month
- 5
Inference Providers
NEW
This model isn't deployed by any Inference Provider.
π
Ask for provider support