CheXNet implementation in PyTorch
Yet another PyTorch implementation of the CheXNet algorithm for pathology detection in frontal chest X-ray images. This implementation is based on approach presented here. Ten-crops technique is used to transform images at the testing stage to get better accuracy.
The highest accuracy evaluated with AUROC was 0.8508 (see the model m-25012018-123527 in the models directory). The same training (70%), validation (10%) and testing (20%) datasets were used as in this implementation.
Prerequisites
- Python 3.5.2
- Pytorch
- OpenCV (for generating CAMs)
Usage
- Download the ChestX-ray14 database from here
- Unpack archives in separate directories (e.g. images_001.tar.gz into images_001)
- Run python Main.py to run test using the pre-trained model (m-25012018-123527)
- Use the runTrain() function in the Main.py to train a model from scratch
This implementation allows to conduct experiments with 3 different densenet architectures: densenet-121, densenet-169 and densenet-201.
- To generate CAM of a test file run script HeatmapGenerator
Results
The highest accuracy 0.8508 was achieved by the model m-25012018-123527 (see the models directory).
| Pathology | AUROC |
|---|---|
| Atelectasis | 0.8321 |
| Cardiomegaly | 0.9107 |
| Effusion | 0.8860 |
| Infiltration | 0.7145 |
| Mass | 0.8653 |
| Nodule | 0.8037 |
| Pneumonia | 0.7655 |
| Pneumothorax | 0.8857 |
| Consolidation | 0.8157 |
| Edema | 0.9017 |
| Emphysema | 0.9422 |
| Fibrosis | 0.8523 |
| P.T. | 0.7948 |
| Hernia | 0.9416 |
Computation time
The training was done using single Tesla P100 GPU and took approximately 22h.