# LungLens

> Chest X-ray TB screening (healthy / sick but not TB / TB) with a heatmap, running in the browser. Deep Learning semester project by Zainab Hamid.

- Live demo: https://lunglens-tb.vercel.app
- Code, weights, data splits: https://github.com/zyna-b/lunglens
- Dataset: TBX11K, https://www.kaggle.com/datasets/vbookshelf/tbx11k-simplified (CC BY 4.0)
- Report: https://lunglens-tb.vercel.app/report.pdf
- Slides: https://lunglens-tb.vercel.app/slides.pdf
- Video: https://www.youtube.com/watch?v=caT1J9Lhu4A
- LinkedIn: https://www.linkedin.com/feed/update/urn:li:activity:7522041517444866048/
- Machine-readable summary: /project.json ; model card: /model_card.json

## Model
- Deployed: DenseNet-121 (ImageNet start); input [1, 1, 384, 384] float32 in [0,1] (grayscale, 384x384); outputs: probs [1,3] (softmax over healthy, sick_but_not_tb, tb) and cam [1,3,12,12].
- Weights: weights/lunglens.onnx (ONNX), weights/lunglens_torchscript.pt (TorchScript), weights/densenet121_imagenet.pt (PyTorch state_dict).
- Screening rule: refer for a lab test if probs[tb] >= 0.0070 (set at 95% specificity on the validation non-TB X-rays).

## Test results (1,800 held-out X-rays)
- Accuracy 99.2%, balanced accuracy 98.4%, TB AUC 0.9996, sensitivity 100.0%, specificity 95.9%.
- Pointing game (heatmap peak inside a radiologist's TB box): 36.0% vs 9.8% for a random point.

## How to test
git clone https://github.com/zyna-b/lunglens && cd lunglens && pip install -r requirements.txt
python evaluate.py                         # 12 bundled labelled X-rays
python evaluate.py --data /path/to/tbx11k  # full test split from data/splits.csv
python predict.py samples/tb_1.png --heatmaps out/