NIH Chest X-ray WebDataset Subset

This dataset is a sample subset of the NIH Chest X-ray Dataset. While its size of 1 GB is only 2.4% of the size of the original dataset, it still allows creating an accurate classifier using the Augmented Chest X-Ray repository. The subset is stored in a collection of WebDataset tar archives for sequential data access. It is ready to be used for training by using the train_model.py module in Augmented Chest X-Ray.

The NIH Chest X-ray Dataset

The NIH Chest X-ray Dataset was introduced in ChestX-ray8 and later ChestX-ray14 by Wang et al.¹ It consists of 112,120 frontal view X-ray images of 30,805 unique patients with 14 common disease labels mined using Natural Language Processing.

The dataset is provided by the NIH Clinical Center and can be downloaded at this link.

The Subset

The subset consists of 30,545 X-ray images of 8,400 unique patients. The images were downscaled to 256x256 pixels and split into training, validation and test sets, ensuring no patient overlap between sets.

The subset was created using the preprocess_data.py module from the Augmented Chest X-Ray repository:

$ python preprocess_data.py -s 8400

Accuracy of a Model Trained on the Subset

A model was trained on the training set for 8 epochs and reached a validation loss of 0.1521. The model was then tested on the provided test set and achieved an AUROC of 0.816.

To better quantify the achieved accuracy, the model was additionally tested on a separate, larger test set (not included in this repository). The larger test set consisted of 27,161 images from 7,500 patients. There was no patient overlap between the subset used for training and either of the test sets. The model achieve an AUROC of 0.809 on the larger test set.

Training and validation was performed using the train_model.py and validate_model.py modules from the Augmented Chest X-Ray repository.

Training the model:

$ python train_model.py -b 16 -w 2 -e 8 -l 0.0001 -a

Validation on the test set:

$ python validate_model.py -b 32 -w 2 -c

Reference

1. ^ Xiaosong Wang, Yifan Peng, Le Lu, Zhiyong Lu, Mohammadhadi Bagheri, Ronald Summers, ChestX-ray8: Hospital-scale Chest X-ray Database and Benchmarks on Weakly-Supervised Classification and Localization of Common Thorax Diseases, IEEE CVPR, pp. 3462-3471, 2017