timm/nih-chest-xray-14

Description

The NIH ChestX-ray14 dataset (NIH Clinical Center), an extension of ChestX-ray8 from the CVPR 2017 paper. It has 112,120 frontal-view chest X-rays of 30,805 unique patients. Each image is labelled with any of 14 thoracic findings, text-mined from the associated radiology reports with NLP.

' ChestX-ray dataset comprises 112,120 frontal-view X-ray images of 30,805 unique patients with the text-mined fourteen disease image labels (where each image can have multi-labels), mined from the associated radiological reports using natural language processing. Fourteen common thoracic pathologies include Atelectasis, Consolidation, Infiltration, Pneumothorax, Edema, Emphysema, Fibrosis, Effusion, Pneumonia, Pleural_thickening, Cardiomegaly, Nodule, Mass and Hernia, which is an extension of the 8 common disease patterns listed in our CVPR 2017 paper. '

Labels (14): Atelectasis, Cardiomegaly, Consolidation, Edema, Effusion, Emphysema, Fibrosis, Hernia, Infiltration, Mass, Nodule, Pleural_Thickening, Pneumonia, Pneumothorax.

Images labelled No Finding in the source (about 54% of images) have an empty labels list. Per the NIH FAQ, 'No Finding' is not the same as 'normal': those images may show patterns outside the 14 categories, or uncertain findings.

The labels are NLP-extracted and noisy (NIH estimates >90% accuracy). The test labels come from the same NLP process and are not radiologist-verified. Improved label sets are linked from the original download site (e.g. MAPLEZ, NIH-CXR-LT).

The splits are the official patient-level train_val_list.txt (86,524 images, 28,008 patients) and test_list.txt (25,596 images, 2,797 patients). No patient appears in both. There is no official validation split. The train split has a fold column (0-9) that I assigned with multi-label iterative stratification (Sechidis et al., 2011, seed 42), grouped by patient: every image of a patient is in the same fold. Use a fold as validation (e.g. fold == 0, ~10%) without leaking patients between train and validation. Findings and 'No Finding' are balanced to within 1% across folds; Hernia, the rarest, has 14-15 per fold.

Images are the original 1024x1024 8-bit PNGs re-encoded as 8-bit grayscale (L mode) JPEG at quality 95. This halves the size relative to PNG (mean PSNR ~48.4 dB, SSIM ~0.988 vs. the PNGs), and all 256 gray levels are preserved. About 1% of the source PNGs are stored as RGBA with identical R/G/B channels and opaque alpha; these were converted to L losslessly before encoding.

• —Website: https://nihcc.app.box.com/v/ChestXray-NIHCC
• —Paper: https://arxiv.org/abs/1705.02315

Dataset Structure

• —image - JPEG image, 1024x1024, 8-bit grayscale.
• —labels - sequence of ClassLabel indices of the positive findings (empty for 'No Finding').
• —label_names - names of the positive findings.
• —image_id - original image filename stem (e.g. 00000001_000).
• —fold - patient-grouped stratified fold (0-9) for train, null for test.
• —patient_id, follow_up, patient_age, patient_sex, view_position (PA / AP) - from Data_Entry_2017_v2020.csv (Patient ID, Follow-up #, Patient Age, Patient Sex, View Position). Per NIH, the image name suffix does not necessarily match the follow-up number.
• —original_width, original_height, pixel_spacing_x, pixel_spacing_y - original image size and pixel spacing before the 1024x1024 resize, from the same file.
• —bboxes - list of {bbox: [x, y, w, h], label} from BBox_List_2017.csv, in pixels on the 1024x1024 image, (x, y) = top-left corner. Present for 880 test images (984 boxes, 8 findings), empty otherwise. The source's Infiltrate is mapped to Infiltration.

Usage

This dataset follows the same multi-label format as timm/plant-pathology-2021 and timm/amazon-from-space for use with timm 1.0.31 or greater. The images are grayscale; timm converts them to 3-channel RGB by default.

python train.py --dataset hfds/timm/nih-chest-xray-14 --train-split train --val-split test \
  --task multilabel --target-key labels --num-classes 14 \
  --model resnet50.ram_in1k --pretrained --img-size 512 --batch-size 32 --epochs 30 \
  --opt adamw --lr 1e-4 --weight-decay 0.01 --amp

To hold out a patient-level validation fold instead of validating on test:

from datasets import load_dataset
ds = load_dataset('timm/nih-chest-xray-14', split='train')
train_ds = ds.filter(lambda f: f != 0, input_columns='fold')
val_ds = ds.filter(lambda f: f == 0, input_columns='fold')

The standard metric is per-finding ROC AUC, averaged over the 14 findings.

License / Terms

From the NIH FAQ: "The usage of the data set is unrestricted. But you should provide the link to our original download site, acknowledge the NIH Clinical Center and provide a citation to our CVPR 2017 paper."

Original download site: https://nihcc.app.box.com/v/ChestXray-NIHCC. Data provided by the NIH Clinical Center.

Citation

@InProceedings{wang2017chestxray,
  author = {Wang, Xiaosong and Peng, Yifan and Lu, Le and Lu, Zhiyong and Bagheri, Mohammadhadi and Summers, Ronald},
  title = {ChestX-ray8: Hospital-scale Chest X-ray Database and Benchmarks on Weakly-Supervised Classification and Localization of Common Thorax Diseases},
  booktitle = {2017 IEEE Conference on Computer Vision and Pattern Recognition (CVPR)},
  pages = {3462--3471},
  year = {2017}
}