Jetson X-ray Pneumonia Classifier

Robotics & Edge Computing Jetson Systems Jetson Projects
,
1

I built a Jetson-based demo that classifies chest X-ray images into NORMAL vs PNEUMONIA.

  • Device: Jetson Nano (JetPack 4.6 / L4T r32.7.1)
  • Data: Kaggle chest radiograph images (pneumonia & normal)
    Source: Chest Radiograph Images (Pneumonia & Normal) | Kaggle
  • Model: ResNet-18 (ImageNet pretrained → fine-tuned to 2 classes)
  • Acceleration: PyTorch → torch2trt → TensorRT (FP16)
  • UI: Flask web app for uploading an X-ray and getting prediction, probabilities, and inference time.

Repo: jetson-xray-pneumonia

Run:

sudo docker run --runtime nvidia -it --rm --network host \
  --shm-size=1g \
  -v /home/jongbum/nvdli-data:/nvdli-data \
  -v /home/jongbum/jetson-xray-pneumonia:/workspace \
  jetson-xray:pneu

cd /workspace
# 1) split dataset (train / val / test)
python3 scripts/split_dataset.py --config configs/default.yaml

# 2) train model on Jetson
python3 scripts/train.py --config configs/default.yaml

# 3) evaluate
python3 scripts/evaluate.py configs/default.yaml

# 4) convert trained model to TensorRT (FP16)
python3 scripts/convert_trt.py configs/default.yaml

# 5) start web app
PYTHONPATH=/workspace/scripts python3 app/app.py
# open in browser: http://<jetson-ip>:5000