Reproduce 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 (with original data set,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.
Result for all models show in result.pdf.
Many thanks to zoogzog for detailed implematation of tranmsformatian dataset and auroc comoutation.
- Python 3.6
- Pytorch==0.3.1
- 2 GPU(recommend NVIDIA TESLA P100 or equivalent)
- OpenCV (for generating CAMs)
- Linux OS(Google cloud platform,debian 9,8CPU(56G RAM)
- Download the ChestX-ray14 database from here
- Unpack archives in separate directories (e.g. images_001.tar.gz into images_001)
- The original dataset contains over 100,000 images,which takes too long to train a model.Thus,we downsize the data with the policy on :drop duplicate where patient id and label are the same from Data_Entry_2017.csv .The smaller dataset contains almost half of original(train_1.txt,test_1.txt,val_1.txt are the original split txt index for immages,while train.txt,test.txt,val.txt are the final splitting txt file to train our model,which took much less time).
The distribution of pathology of original and downsize data show below,which doesn't change much:
- Run python Main.py to run test using the pre-trained model (m-25012018-123527 for DenseNet121)
- Use the runTrain() function in the Main.py to train a model from scratch
This implementation allows to conduct experiments with 4 different densenet architectures: densenet-121, densenet-169,ResNet50,Se_ResNet50,Se_DenseNet121.
- To generate CAM of a test file run script HeatmapGenerator
The highest accuracy 0.8508 was achieved by the model m-25012018-123527 for DenseNet121 started by pretrained model(see the models directory).
| Pathology | AUROC |
|---|---|
| Mean | 0.8508 |
| 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 |
The training was done using double Tesla P100 GPU and 6h on downsize data.
CheXNet,Squeeze-and-Excitation Networks, DenseNet and ResNet.