LungGuardian

Pneumonia detection from chest X-rays using Faster R-CNN with EfficientNetV2 backbone. Built for the RSNA Pneumonia Detection Challenge.

What it does

This is an object detection model, not a classifier. It finds and localizes pneumonia opacities in chest X-rays with bounding boxes.

Input	Output

Tech stack

Component	Choice	Why
Backbone	EfficientNetV2-S	Better accuracy/speed than ResNet
Detector	Faster R-CNN + FPN	Multi-scale detection
Loss	Focal Loss	Handles 95%+ background anchors
Post-processing	Weighted Box Fusion	Better than NMS for overlapping boxes
Augmentation	Albumentations + Mosaic	Robust to variations
CV Strategy	StratifiedGroupKFold	No patient leakage between folds

Setup

git clone https://github.com/yourusername/LungGuardian.git
cd LungGuardian
pip install -r requirements.txt

Download the RSNA dataset and extract to data/rsna/.

Requirements

• Python 3.8+
• PyTorch 2.0+
• CUDA 11.0+ (for GPU)
• ~8GB VRAM

Training

# Single fold
python notebooks/04_training_engine.py --fold 0 --epochs 15

# All 5 folds
python notebooks/04_training_engine.py --all_folds

# Resume from checkpoint
python notebooks/04_training_engine.py --fold 0 --resume checkpoints/run_xxx/best_model.pt

Project structure

├── notebooks/
│   ├── 01_data_pipeline.py        # DICOM loading, augmentation, StratifiedGroupKFold
│   ├── 02_model_architecture.py   # EfficientNetV2 + FPN + Faster R-CNN
│   ├── 03_losses_and_utilities.py # Focal loss, WBF, IoU, mAP
│   └── 04_training_engine.py      # Training loop with AMP, scheduling
├── assets/                        # Demo images
├── checkpoints/                   # Saved models (gitignored)
├── requirements.txt
└── README.md

Config

Param	Value	Notes
Image size	1024×1024	Native RSNA resolution
Batch size	4	Gradient accumulation = 4 (effective 16)
LR backbone	1e-5	Lower LR for pretrained layers
LR head	1e-4	Higher LR for new layers
Scheduler	Cosine + warmup	2 epoch warmup
Epochs	15	Early stopping with patience=5
AMP	Enabled	~2x faster, 50% less memory

Data pipeline

The RSNA dataset has some quirks this code handles:

• Multiple bounding boxes per image (grouped by patientId)
• ~26% positive rate (handled via Focal Loss)
• Same patient can't be in train AND val (StratifiedGroupKFold)

Augmentations: HorizontalFlip, ShiftScaleRotate, RandomBrightnessContrast, CLAHE, GaussNoise, Mosaic (30% prob)

Inference

model = create_efficientnetv2_s_fasterrcnn(num_classes=2)
model.load_state_dict(torch.load('best_model.pt')['model_state_dict'])
model.eval()

with torch.no_grad():
    preds = model([image_tensor])

# Apply WBF to merge overlapping boxes
preds = apply_wbf(preds, image_size=1024, iou_thr=0.5)

References

• EfficientNetV2 - Tan & Le, 2021
• Faster R-CNN - Ren et al., 2015
• Feature Pyramid Networks - Lin et al., 2017
• Focal Loss - Lin et al., 2017
• Weighted Box Fusion - Solovyev et al., 2021

License

MIT