Technical Breakdown
Pneumonia Detection
26,684 RSNA chest X-rays (70% train, 15% val/test). Binary classification. 3.44:1 class imbalance.
Approach
Swin Transformer for pneumonia detection from RSNA chest X-rays. Hierarchical attention captures local and global context.
Objective
Build pneumonia detector for medical screening with high sensitivity.
Challenge
77% normal vs 23% pneumonia cases. Manual radiologist review slow and inconsistent.
Solution
Swin-T + WeightedRandomSampler for imbalance. ImageNet pretraining, Adam LR=1e-4.
Dataset
RSNA (26,684)
Type
Binary
Model
Swin-T
Functions
- DICOM → image conversion, resize, normalize.
- WeightedRandomSampler for class balance.
- Swin-T with ImageNet pretraining.
- Accuracy, precision, recall, F1, ROC-AUC.
Architecture
- Swin-T backbone for hierarchical features.
- Shifted window attention, O(n) complexity.
- ImageNet transfer learning.
- Sigmoid for binary output.
- Binary cross-entropy + class weights.
Preprocessing
- DICOM conversion.
- Resize 224×224.
- ImageNet normalization.
- No augmentation (preserve diagnostics).
- Weighted batch sampling.
Training
- Loss: BCE with weights (normal 0.22, pneumonia 0.78).
- Adam, LR=1e-4, batch=16.
- 10 epochs, early stopping.
- Metrics: accuracy, precision, recall, F1.
Binary Classification Performance (Swin Transformer Test Results)
| Classification Class | Precision | Recall | F1-Score |
|---|---|---|---|
| Normal (No Pneumonia) | 0.97 | 0.66 | 0.78 |
| Pneumonia (Positive) | 0.44 | 0.92 | 0.59 |
Test accuracy: 71.52% overall. Model prioritizes recall (92.13%) for pneumonia detection to minimize missed diagnoses. Precision-recall trade-off is acceptable in medical screening where sensitivity is critical. Weighted F1-score: 0.74.