Towards Semantic Interpretation of Thoracic Disease and COVID-19 Diagnosis Models
This is the source code for our paper. The contributions of this code are two-fold:
- Code for training and evaluation of COVID-19 models, mentioned in the paper.
- Network dissection setup for analysing the trained models.
Requirements
- pytorch 1.7
- torchvision 0.8.1
- tensorboard 2.4.0
- opencv 4.0.1
- scikit-learn 0.23.2
Training and evaluation
Data
The datasets CheXpert, BrixIA, NIH ChestXray, have to be manually downloaded and put in the data folder. This is because for CheXpert and BrixIA it is required to register to gain access to the data. Under /src/data/ two preprocessing functions are provided one for BrixIA and one for CheXpert.
-
preprocess_brixia.py :
- Converts .dcm images to .jpg
- Transforms images with torchvision transforms
-
preprocess_chexpert.py :
- Transforms images with torchvision transforms
Model training & evaluation
Functionalities to train and evaluate a model (DenseNet121) are in the main.py. The model is trained under a given config and afterwards evaluated. The model weights and TensorBoard log files are stored in the runs folder.
CheXpert model
A function to train and evaluate a model on the CheXpert dataset on 14 pathologies with either weighted BCE or unweighted BCE is provided.
chexpert_model(pretrained_model_path=None, **train_config)Training parameters and experiment name can be defined in the train_config, and a pretrained model can also be defined. If no pretrained model is given, the pretrained weights on ImageNet are taken.
pretrained_model_path = "./runs/chexpert_pretrained/model.pth" # Path to to a pretrained model.pth (DenseNet121) otherwise ImageNet weights are used
train_config = {
'batch_size': 60, # Number of images per batch
'input_size': (224, 224), # Image size of the model input
'n_epochs': 20, # Number of max epochs
'orientation': 'frontal', # Only use the frontal images of CheXpert ('lateral', 'frontal', 'all)
'optim': torch.optim.Adam, # PyTorch optimizer
'optim_kwargs': {'lr': 0.001, 'weight_decay': 0.0}, # Optimizer parameters
'scheduler': torch.optim.lr_scheduler.ReduceLROnPlateau, # PyTorch scheduler
'scheduler_kwargs': {'factor': 0.1, 'patience': 3, 'mode': 'max'}, # Scheduler parameters
'weighted_bce': False, # Use weighted or unweighted BCE
'early_stopping': 10, # Stop training after n epochs if AUC does not improve
'experiment_name': 'chexpert_model', # Experiment name for runs folder
}BrixIA model
A function to train a model on the BrixIA dataset on single target regression and six zone classification/regression.
brixia_model(pretrained_model_path=None, **train_config)Similar to before the training parameters can be specified in the train_config. A previously trained CheXpert model can be used as pretrained model.
pretrained_model_path = "./runs/chexpert_pretrained/model.pth" # Path to to a pretrained model.pth (DenseNet121) otherwise ImageNet weights are used
train_config = {
'batch_size': 60, # Number of images per batch
'input_size': (224, 224), # Image size of the model input
'n_epochs': 20, # Number of max epochs
'optim': torch.optim.Adam, # PyTorch optimizer
'optim_kwargs': {'lr': 0.001, 'weight_decay': 0.0}, # Optimizer parameters
'scheduler': torch.optim.lr_scheduler.ReduceLROnPlateau, # PyTorch scheduler
'scheduler_kwargs': {'factor': 0.1, 'patience': 3, 'mode': 'max'}, # Scheduler parameters
'early_stopping': 6, # Stop training after n epochs if AUC does not improve
'experiment_name': 'brixia_model', # Experiment name for runs folder
'alpha': 0.7, # If mode=paper then alpha gives a balancing between the two loss functions for the six zone training
'mode': 'paper', # 'paper' for six zone model, 'regression' for single target model
'criterion': torch.nn.MSELoss(), # If mode=regression the criterion specified here is used
}Combined model
A function to train a combined model on binary classification (Finding/NoFinding) on three datasets CheXpert, NIH ChestXray and BrixIA. The train_config is similar to the CheXpert config.
combined_model(pretrained_model_path=None, **train_config)Network Dissection
You can analyze any of the trained models with network dissection to identify concepts that the network is looking for.
Data
We used two datast to interpret the network results. You would need to download these datasets first, before running any network dissection analysis.
Interpreting a model
You can see a sample analysis with network dissection of the BrixIA model in Covid_seg_network_dissection_results notebook You can replace the model defined there with a model of your choosing and run the network dissection analysis. Please refer to the explanations in the notebook to understand the resulst.