ChestX-Det

3,578 frontal chest radiographs from NIH ChestX-ray14, annotated by three board-certified radiologists with 13 categories of thoracic abnormality — each as a per-instance polygon contour plus a bounding box.

• Modality: chest X-ray (2D frontal radiograph), 1024×1024, 8-bit
• Organ: thorax — lung, pleura, mediastinum, ribs, heart
• Targets: 13 abnormality classes, as independent binary masks
• License: Apache-2.0 (annotations) over NIH ChestX-ray14 images (no use restrictions, attribution required)
• Source: Deepwise-AILab/ChestX-Det-Dataset (official)

⚠️ ChestX-Det vs ChestX-Det10 — only one is a segmentation dataset

	ChestX-Det (this repo)	ChestX-Det10
Images	3,578	3,543 (exact subset)
Classes	13	10
polygons key	present in all 3,578	absent in all 3,543
License	Apache-2.0	none — all rights reserved

ChestX-Det10 is boxes-only and cannot be used for segmentation. Much of the literature writes "ChestX-Det" when it means Det10 — check for the polygons key. ChestX-Det adds 3 classes (Cardiomegaly, Diffuse Nodule, Pleural Thickening), 35 images, and the polygon contours.

Splits

Both official. There is no validation split.

Split	Images	Instances	No-finding images
train	3,025	7,905	547
test	553	1,734	64
Total	3,578	9,639	611 (17.1%)

Classes

Counts are (image, class) pairs — i.e. images on which that class appears — followed by instance counts.

Class	Column prefix	Images	Instances
Consolidation	consolidation	1,642	2,563
Effusion	effusion	1,441	2,112
Fibrosis	fibrosis	500	739
Pleural Thickening	pleural_thickening	499	631
Nodule	nodule	392	986
Fracture	fracture	389	661
Atelectasis	atelectasis	297	340
Cardiomegaly	cardiomegaly	293	293
Calcification	calcification	202	348
Pneumothorax	pneumothorax	189	211
Emphysema	emphysema	176	298
Diffuse Nodule	diffuse_nodule	169	290
Mass	mass	148	167

Columns

Column	Type	Notes
image_id	string	Upstream stem, e.g. 36199 — the pairing key
image	Image (L)	Radiograph, 1024×1024. Normalised to mode L — see below
<class>_mask ×13	Image (L), nullable	{0,255}, 255 = abnormality. Union of that class's instances
has_<class>_mask ×13	bool	Explicit presence flag
<class>_fraction ×13	float32, nullable	Foreground area fraction of the frame
annotation_json	string	Original upstream entry verbatim — syms, boxes, per-instance polygons
num_instances	int32	Total annotated instances on this image
num_classes_present	int32	Distinct classes on this image (0–13)
is_negative	bool	True for the 611 images with no finding
width, height	int32	Always 1024×1024
original_mode	string	L or RGBA before normalisation

⚠️ The 13 classes overlap at the pixel level

The ChestX-Det10 paper states it outright: "multiple diseases might share the same region of box." Measured by rasterising all 9,639 polygons:

• 2,006 / 3,578 images carry more than one class
• 1,357 of those 2,006 (67.6 %) have genuine pixel overlap between classes
• mean IoU 0.272 where overlap occurs
• Consolidation∩Effusion overlap in 770 of their 1,028 co-occurring images; Cardiomegaly∩Consolidation in 148 of 158

They are therefore 13 independent binary masks, not one label map. Collapsing them into a single {0..13} image would silently erase whichever class is composited first.

⚠️ An absent class is negative, not unlabelled

The radiologists annotated all 13 categories on every image, so a class with no polygon is genuinely absent from that radiograph. Absence is stored as a null mask plus a has_<class>_mask boolean — never infer it from an empty mask. 86.4 % of the 13 × 3,578 cells are null.

Separately, 611 images (17.1 %) have no finding at all (is_negative), and 22.1 % of the annotated (image, class) pairs cover < 0.5 % of the frame (median 2.5 %). Any small-n debug sample will be badly unrepresentative.

⚠️ Image mode was normalised here

42 of the 3,578 upstream PNGs (37 train / 5 test) are mode RGBA; the other 3,536 are mode L — a quirk inherited from NIH ChestX-ray14. In all 42, R == G == B exactly (max inter-channel difference 0) and alpha is uniformly 255, so they are grayscale wearing four channels and .convert("L") is bit-for-bit lossless. Every image here is mode L. original_mode records what it was upstream.

⚠️ No patient ID, and the NIH filename is destroyed

Deepwise renamed every file to an opaque integer (36199 … 70958). The NIH ChestX-ray14 filename (00000013_005.png) is not recoverable and no mapping is published. Consequences:

• No patient grouping is possible. NIH averages ~3.6 images per patient and the official split is described as random over images, so intra-dataset patient leakage between train and test is likely and cannot be repaired.
• No cross-dataset de-duplication is possible. Every image is a member of NIH ChestX-ray14 by construction. SIIM-ACR Pneumothorax is the live hazard — it is built on the same NIH source, so overlap with this dataset's Pneumothorax class is plausible and unquantifiable. Do not treat the two as independent benchmarks.
• No overlap with CheXpert, MIMIC-CXR, JSRT, Montgomery/Shenzhen, COVID-19 Radiography or COVID-QU-Ex (all different source institutions).

Ground truth — one adjudicated tier

Three board-certified radiologists in a 2+1 protocol. Two annotate every image independently and mutually blind (the committee); a third with 15+ years' experience adjudicates, selecting among their annotations and adding any they missed (the judge). Both papers designate the judge's output the gold standard for train and test alike, while explicitly noting residual subjective variation.

Only the adjudicated result is released — the two committee members' individual annotations are not — so there is no inter-rater agreement ceiling for this dataset and no rater or tier to choose between.

Paper vs release: trust the release

The TMI paper's Table 1 disagrees slightly with the shipped JSON: Nodule 848/182 (paper) vs 816/170 (release) and Diffuse Nodule 195/51 vs 227/63 — a clean transfer of exactly 32 train + 12 test instances between the two nodule classes, i.e. a post-publication reclassification. Also Fracture 547→546 and Effusion 379→378. Paper total 9,641 vs released 9,639. The paper text says "3,575 images" while 3,025 + 553 = 3,578 are released, matching the README. The counts in this card are re-derived from the released JSONs.

Provenance

Fetched from the official Deepwise host and GitHub repository — not a mirror. Both zips match their Content-Length byte-exactly; every JSON filename maps 1:1 to a file on disk in both directions; all images are 1024×1024; instance totals reconcile with the upstream README exactly. Three unofficial HF mirrors of this dataset exist (Lin-Mars/ChestX_Det, natealberti/ChestX-Det) and none was used. Note that the Kaggle mirror mathurinache/chestxdetdataset tags this CC BY-NC-SA 4.0, which contradicts the upstream Apache-2.0 — an uploader-assigned label, not the primary source.

Attribution (required)

Per NIH terms, use of the underlying images requires all three of:

1. A link to the NIH download page: https://nihcc.app.box.com/v/ChestXray-NIHCC
2. Citation of Wang et al., CVPR 2017 (below)
3. Acknowledgement that the NIH Clinical Center is the data provider

Citation

@article{lian2021sarnet,
  title   = {A Structure-Aware Relation Network for Thoracic Diseases Detection
             and Segmentation},
  author  = {Lian, Jie and Liu, Jingyu and Zhang, Shu and Gao, Kai and
             Liu, Xiaoqing and Zhang, Dingwen and Yu, Yizhou},
  journal = {IEEE Transactions on Medical Imaging},
  year    = {2021},
  eprint  = {2104.10326}
}

@inproceedings{liu2020chestxdet10,
  title     = {ChestX-Det10: Chest X-ray Dataset on Detection of Thoracic
               Abnormalities},
  author    = {Liu, Jingyu and Lian, Jie and Yu, Yizhou},
  booktitle = {arXiv:2006.10550},
  year      = {2020}
}

@inproceedings{wang2017chestxray8,
  title     = {ChestX-ray8: Hospital-scale Chest X-ray Database and Benchmarks on
               Weakly-Supervised Classification and Localization of Common Thorax
               Diseases},
  author    = {Wang, Xiaosong and Peng, Yifan and Lu, Le and Lu, Zhiyong and
               Bagheri, Mohammadhadi and Summers, Ronald M.},
  booktitle = {IEEE CVPR},
  pages     = {3462--3471},
  year      = {2017}
}