In this page, interpretation of lesional detection via counterfactual generation framework is devised to provide visual interpretation for classifying chest X-ray images.
- python 3.7.11
- numpy 1.20.3
- h5py 3.4.0
- pytorch 1.6.0
- pillow 8.3.1
- scipy 1.6.2
- torchray 1.0.0.2
- torchvision 0.7.0
- opencv-python 4.5.3.56
Due to upload limit of Github, this page does not include checkpoint files. Please send request email to [email protected] if the checkpoint files are needed. Once you have the parameters, create two folders './parameters/CheXNet_ChestX-ray14/' and 'parameters/GAN_ChestX-ray14/' then locate each parameter in it.
Here we provide the inference model for Chest X-ray 14 dataset lesional detection.
We include two samples of the dataset in './dataset/image' folder, and a text file including directory of the samples, class and bounding box information in './dataset/list'. The format of the dataloader text file is as follows,
'00017544_003.png 0 0 0 1 0 1 0 0 0 1 0 0 0 0 585.386666666667 178.953489583333 288.995555555556 584.817777777778 0 0 0 1 0 0 0 0 0 0 0 0 0 0',
where the first component is the file name of the image, the second is multi-label ground truth one-hot vector, the third is the bounding box coordinate in (x,y,w,h) order, and the last component is the ground truth one-hot vector for the included bounding box.
For further inference with the full-dataset, please infer to 'https://www.kaggle.com/nih-chest-xrays/data' to download the dataset. The images are expected to be included in './dataset/' folder, however, the users can modify the dataset directory in main.py file by modifying the argument parser parameter.
By running the main.py, the lesional detection results are to be created in './outputs' folder. The outputs will be three for one image, where 'Original_Image_#' is the original chest X-ray image, 'Generated_image_#' is the generated image by GAN model, and 'CheXGAN_Visualization_LesionName_#' is the difference map between the original image and the generated image.