torchxrayvision

A library for chest X-ray datasets and models. Including pre-trainined models.

(🎬 promo video about the project)

Motivation: While there are many publications focusing on the prediction of radiological and clinical findings from chest X-ray images much of this work is inaccessible to other researchers.

• In the case of researchers addressing clinical questions it is a waste of time for them to train models from scratch. To address this, TorchXRayVision provides pre-trained models which are trained on large cohorts of data and enables 1) rapid analysis of large datasets 2) feature reuse for few-shot learning.
• In the case of researchers developing algorithms it is important to robustly evaluate models using multiple external datasets. Metadata associated with each dataset can vary greatly which makes it difficult to apply methods to multiple datasets. TorchXRayVision provides access to many datasets in a uniform way so that they can be swapped out with a single line of code. These datasets can also be merged and filtered to construct specific distributional shifts for studying generalization.

This code is still under development

Twitter: @torchxrayvision

Getting started

pip install torchxrayvision

import torchxrayvision as xrv

These are default pathologies:

xrv.datasets.default_pathologies 

['Atelectasis',
 'Consolidation',
 'Infiltration',
 'Pneumothorax',
 'Edema',
 'Emphysema',
 'Fibrosis',
 'Effusion',
 'Pneumonia',
 'Pleural_Thickening',
 'Cardiomegaly',
 'Nodule',
 'Mass',
 'Hernia',
 'Lung Lesion',
 'Fracture',
 'Lung Opacity',
 'Enlarged Cardiomediastinum']

Models (demo notebook)

Specify weights for pretrained models (currently all DenseNet121) Note: Each pretrained model has 18 outputs. The all model has every output trained. However, for the other weights some targets are not trained and will predict randomly becuase they do not exist in the training dataset. The only valid outputs are listed in the field {dataset}.pathologies on the dataset that corresponds to the weights.

model = xrv.models.DenseNet(weights="all")
model = xrv.models.DenseNet(weights="rsna") # RSNA Pneumonia Challenge
model = xrv.models.DenseNet(weights="nih") # NIH chest X-ray8
model = xrv.models.DenseNet(weights="pc") # PadChest (University of Alicante)
model = xrv.models.DenseNet(weights="chex") # CheXpert (Stanford)
model = xrv.models.DenseNet(weights="mimic_nb") # MIMIC-CXR (MIT)
model = xrv.models.DenseNet(weights="mimic_ch") # MIMIC-CXR (MIT)

# DenseNet121 from JF Healthcare for the CheXpert competition
model = xrv.baselinemodels.jfhealthcare.DenseNet() 

The performance of these models can be seen here per task as computed in arxiv.org/abs/2002.02497 in terms of AUC. All valid combinations of model and dataset are computed where a model was trained on the specific label and that label exists in the target dataset. A white cell means it cannot be computed due to missing labels in train or test dataset. The outputs of 3 models are averaged together to reduce noise. Each of the 3 models is trained on the same data with different weight initialization. The XRV pretrained models are only one of the seeds.

Autoencoders

You can also load a pre-trained autoencoder that is trained on the PadChest, NIH, CheXpert, and MIMIC datasets.

ae = xrv.autoencoders.ResNetAE(weights="101-elastic")
z = ae.encode(image)
image2 = ae.decode(z)

Datasets (demo notebook)

Only stats for PA/AP views are shown. Datasets may include more.

transform = torchvision.transforms.Compose([xrv.datasets.XRayCenterCrop(),
                                            xrv.datasets.XRayResizer(224)])

d_kaggle = xrv.datasets.RSNA_Pneumonia_Dataset(imgpath="path to stage_2_train_images_jpg",
                                       transform=transform)
                
d_chex = xrv.datasets.CheX_Dataset(imgpath="path to CheXpert-v1.0-small",
                                   csvpath="path to CheXpert-v1.0-small/train.csv",
                                   transform=transform)

d_nih = xrv.datasets.NIH_Dataset(imgpath="path to NIH images")

d_nih2 = xrv.datasets.NIH_Google_Dataset(imgpath="path to NIH images")

d_pc = xrv.datasets.PC_Dataset(imgpath="path to image folder")


d_covid19 = xrv.datasets.COVID19_Dataset() # specify imgpath and csvpath for the dataset

d_siim = xrv.datasets.SIIM_Pneumothorax_Dataset(imgpath="dicom-images-train/",
                                                csvpath="train-rle.csv")

National Library of Medicine Tuberculosis Datasets paper

d_nlmtb = xrv.datasets.NLMTB_Dataset(imgpath="path to MontgomerySet or ChinaSet_AllFiles")

Using MontgomerySet data:
NLMTB_Dataset num_samples=138 views=['PA']
{'Tuberculosis': {0: 80, 1: 58}}
or using ChinaSet_AllFiles data:
NLMTB_Dataset num_samples=662 views=['PA', 'AP']
{'Tuberculosis': {0: 326, 1: 336}}

Dataset fields

Each dataset contains a number of fields. These fields are maintained when xrv.datasets.Subset_Dataset and xrv.datasets.Merge_Dataset are used.

Each dataset has a .pathologies field which is a list of the pathologies contained in this dataset that will be contained in the .labels field ].

Each dataset has a .labels field which contains a 1,0, or NaN for each label defined in .pathologies.

Each dataset has a .csv field which corresponds to pandas DataFrame of the metadata csv file that comes with the data. Each row aligns with the elements of the dataset so indexing using .iloc will work.

If possible, each dataset's .csv will have some common fields of the csv. These will be aligned when The list is as follows:

csv.patientid A unique id that will uniqely identify samples in this dataset

csv.offset_day_int An integer time offset for the image in the unit of days. This is expected to be for relative times and has no absolute meaning although for some datasets it is the epoch time.

Dataset tools

relabel_dataset will align labels to have the same order as the pathologies argument.

xrv.datasets.relabel_dataset(xrv.datasets.default_pathologies , d_nih) # has side effects

specify a subset of views (demo notebook)

d_kaggle = xrv.datasets.RSNA_Pneumonia_Dataset(imgpath="...",
                                               views=["PA","AP","AP Supine"])

specify only 1 image per patient

d_kaggle = xrv.datasets.RSNA_Pneumonia_Dataset(imgpath="...",
                                               unique_patients=True)

obtain summary statistics per dataset

d_chex = xrv.datasets.CheX_Dataset(imgpath="CheXpert-v1.0-small",
                                   csvpath="CheXpert-v1.0-small/train.csv",
                                 views=["PA","AP"], unique_patients=False)

CheX_Dataset num_samples=191010 views=['PA', 'AP']
{'Atelectasis': {0.0: 17621, 1.0: 29718},
 'Cardiomegaly': {0.0: 22645, 1.0: 23384},
 'Consolidation': {0.0: 30463, 1.0: 12982},
 'Edema': {0.0: 29449, 1.0: 49674},
 'Effusion': {0.0: 34376, 1.0: 76894},
 'Enlarged Cardiomediastinum': {0.0: 26527, 1.0: 9186},
 'Fracture': {0.0: 18111, 1.0: 7434},
 'Lung Lesion': {0.0: 17523, 1.0: 7040},
 'Lung Opacity': {0.0: 20165, 1.0: 94207},
 'Pleural Other': {0.0: 17166, 1.0: 2503},
 'Pneumonia': {0.0: 18105, 1.0: 4674},
 'Pneumothorax': {0.0: 54165, 1.0: 17693},
 'Support Devices': {0.0: 21757, 1.0: 99747}}

Pathology masks (demo notebook)

Masks are available in the following datasets:

xrv.datasets.RSNA_Pneumonia_Dataset() # for Lung Opacity
xrv.datasets.SIIM_Pneumothorax_Dataset() # for Pneumothorax
xrv.datasets.NIH_Dataset() # for Cardiomegaly, Mass, Effusion, ...

Example usage:

d_rsna = xrv.datasets.RSNA_Pneumonia_Dataset(imgpath="stage_2_train_images_jpg", 
                                            views=["PA","AP"],
                                            pathology_masks=True)
                                            
# The has_masks column will let you know if any masks exist for that sample
d_rsna.csv.has_masks.value_counts()
False    20672
True      6012       

# Each sample will have a pathology_masks dictionary where the index 
# of each pathology will correspond to a mask of that pathology (if it exists).
# There may be more than one mask per sample. But only one per pathology.
sample["pathology_masks"][d_rsna.pathologies.index("Lung Opacity")]

it also works with data_augmentation if you pass in data_aug=data_transforms to the dataloader. The random seed is matched to align calls for the image and the mask.

Distribution shift tools (demo notebook)

The class xrv.datasets.CovariateDataset takes two datasets and two arrays representing the labels. The samples will be returned with the desired ratio of images from each site. The goal here is to simulate a covariate shift to make a model focus on an incorrect feature. Then the shift can be reversed in the validation data causing a catastrophic failure in generalization performance.

ratio=0.0 means images from d1 will have a positive label ratio=0.5 means images from d1 will have half of the positive labels ratio=1.0 means images from d1 will have no positive label

With any ratio the number of samples returned will be the same.

d = xrv.datasets.CovariateDataset(d1 = # dataset1 with a specific condition
                                  d1_target = #target label to predict,
                                  d2 = # dataset2 with a specific condition
                                  d2_target = #target label to predict,
                                  mode="train", # train, valid, and test
                                  ratio=0.9)

Citation

Joseph Paul Cohen, Joseph Viviano, Paul Morrison, Rupert Brooks, Mohammad Hashir, Hadrien Bertrand 
TorchXRayVision: A library of chest X-ray datasets and models. 
https://github.com/mlmed/torchxrayvision, 2020

@article{Cohen2020xrv,
author = {Cohen, Joseph Paul and Viviano, Joseph and Morrison, Paul and Brooks, Rupert and Hashir, Mohammad and Bertrand, Hadrien},
journal = {https://github.com/mlmed/torchxrayvision},
title = {{TorchXRayVision: A library of chest X-ray datasets and models}},
url = {https://github.com/mlmed/torchxrayvision},
year = {2020}
}

and this paper https://arxiv.org/abs/2002.02497

Joseph Paul Cohen and Mohammad Hashir and Rupert Brooks and Hadrien Bertrand
On the limits of cross-domain generalization in automated X-ray prediction. 
Medical Imaging with Deep Learning 2020 (Online: https://arxiv.org/abs/2002.02497)

@inproceedings{cohen2020limits,
  title={On the limits of cross-domain generalization in automated X-ray prediction},
  author={Cohen, Joseph Paul and Hashir, Mohammad and Brooks, Rupert and Bertrand, Hadrien},
  booktitle={Medical Imaging with Deep Learning},
  year={2020},
  url={https://arxiv.org/abs/2002.02497}
}

Supporters/Sponsors

We thank CIFAR (Canadian Institute for Advanced Research) for supporting the development of this project.