Pneumothorax Detection & Segmentation
U-Net with EfficientNet-B4 backbone for pneumothorax detection and segmentation on chest X-rays.
Model Details
- Architecture: U-Net with ImageNet-pretrained EfficientNet-B4 encoder
- Input: 512×512 chest X-ray images (3-channel)
- Output: Segmentation mask (512×512) + Classification logit
Training
- Dataset: SIIM-ACR Pneumothorax Segmentation (2,379 positive cases)
- Optimizer: AdamW (lr=1e-4)
- Loss: Focal + Dice (segmentation) + BCE (classification)
- Early stopping: After 7 epochs of no improvement
- Best epoch: 10 / 20 total
Performance (Test Set, n=1,602)
- Classification AUC: 0.9464
- Sensitivity: 81.6%, Specificity: 92.1%
- Segmentation Dice: 0.5142 (lesion-positive cases)
Operating Points
| Threshold | Sensitivity | Specificity |
|---|
| 0.3 | 85.6% | 90.8% |
| 0.5 | 81.6% | 92.1% |
| 0.68 | 80.2% | 93.2% |
Limitations
- Single-dataset training (SIIM-ACR only)
- Sensitivity 2.7 pp below GE predicate at matched specificity
- Segmentation quality (Dice 0.51) indicates room for improvement
- Not validated for clinical use without independent verification
Next Steps
- Multi-dataset training on MIMIC-CXR for better generalization
- Threshold tuning for deployment
Citation
Dataset: SIIM-ACR Pneumothorax Segmentation Challenge (Kaggle 2019)
License
MIT - See accompanying repository for details
Disclaimer
Research model only. Not intended for clinical use without regulatory clearance.
