CheXpert is a large dataset of chest X-rays and competition for automated chest x-ray interpretation
pip install -r requirements.txt
python preprocess_metadata.py
train_processed_0_4.csv CheXpert-v1.0-processed_0_4.zip
- Remove rows without specified value in column Age
- Keep only x-rays with frontal view (AP)
- Treat uncertainty(-1), 50%-50% rule
- Sub-sample full dataset for computational reasons(use only 0.4 of entire set, >60K)
jupyter notebook CheXpert-nb-small.ipynb
We only take 2 approaches here
- U-zeros -> all uncertainty labels treated as no label activation
- U-ones -> all uncertainty labels treated as label activation
- DensetNet121
- EfficientNetB0
- Confusion Matrix
- Receiver Operating Characteristics (ROC) curve
- Gradient-weighted Class Activation Mapping (Grad-CAM)