Pneumonia-Detection

• Analyze chest xrays for diagonising pneumonia and locate the regions with heat map.
• In this work, I used a Transfer Learning Approach to diagnose Pneumonia, by fine-tuning the CheXNet a densenet model developed by Stanford ML Group

Data

Download chestXray data from here to chest_xray folder

Instructions to run

User can run each of the tasks as simple python scripts, with different options as command line arguments.Below are the sample command line executions for each tasks

Training:

 python train.py --batch-size=64 --data_dir='chest-xray/' --lr=0.01

Evaluation:

  python evaluation.py --batch-size=64 --data_dir='chest-xray/' --model_path=runs/debug/chexnet_0.01_5

Testing on Single Image:

python predict.py --model_path=runs/debug/chexnet_0.01_5 --image_path='chest_xrat/test/NORMAL/NORMAL2-IM-0317-0001.jpeg'

Generating Heat Map:

python heatmap.py --model_path=runs/debug/chexnet_0.01_5 --inp_img_pth='chest_xray/test/PNEUMONIA/person59_virus_116.jpeg' --out_img_pth='heatmap.png'

Results and Conclusions:

Densenet, Resnet and CheXnet were experimented for transfer learning, among which CheXnet performed better with 92 % accuracy on test data.

Input Image of person with pneumonia	heatmap

This was further extended to build a web application

References:

• Stanford ML Group

• Reproduce CheXNet

• Data