OpenI Indiana University CXR (IU X-Ray)¶

Modality: CXR | Format: PNG + DICOM | Dim: 2D | Labels: 8 binary findings (CheXpert-compatible)

Overview¶

3,955 radiology reports with 7,470 paired frontal and lateral chest X-ray images from the Indiana University hospital network, released by the NLM/LHNCBC under an open-access licence.

7,470 images — 2 per study (frontal PA + lateral), each stored as both PNG and DICOM
3,955 XML reports — structured with FINDINGS, IMPRESSION, INDICATION sections and MeSH major terms
Labels — 8 binary columns derived from MeSH major terms via prefix matching

Download¶

BASE="https://openi.nlm.nih.gov/imgs/collections"
DEST="./OpenI-IU-CXR"

# PNG images (~1.4 GB)
wget -P "$DEST" "${BASE}/NLMCXR_png.tgz"
tar -xzf "$DEST/NLMCXR_png.tgz" -C "$DEST"
mv "$DEST"/*.png "$DEST/images/"

# Reports (~1.1 MB)
wget -P "$DEST" "${BASE}/NLMCXR_reports.tgz"
tar -xzf "$DEST/NLMCXR_reports.tgz" -C "$DEST"

# DICOM (~75 GB, optional)
wget -P "$DEST" "${BASE}/NLMCXR_dcm.tgz"
tar -xzf "$DEST/NLMCXR_dcm.tgz" -C "$DEST"

Expected layout¶

OpenI-IU-CXR/                 ← base_image_dir points here
  ecgen-radiology/
    1.xml  ...  3955.xml
  images/
    CXR1_1_IM-0001-3001.png
    CXR1_1_IM-0001-4001.png
    ...
  1/                           ← DICOM folders (one per patient, optional)
    1_IM-0001-3001.dcm
    1_IM-0001-4001.dcm
  2/
    2_IM-0652-1001.dcm
    ...

Label columns¶

Derived from <MeSH><major> tags in the XML via case-insensitive prefix matching.

Column	Positive count	MeSH triggers
`no_finding`	2,783	"normal"
`cardiomegaly`	568	"cardiomegaly"
`edema`	118	"pulmonary edema", "pulmonary congestion", "edema"
`atelectasis`	539	"atelectasis", "pulmonary atelectasis", "lung/hypoinflation"
`consolidation`	270	"consolidation", "pneumonia", "airspace disease"
`pleural_effusion`	203	"pleural effusion", "effusion/pleural"
`pneumothorax`	35	"pneumothorax"
`support_devices`	225	"catheters", "pacemaker", "leads", "tube", "defibrillator"

Extra metadata columns¶

Column	Type	Description
`dicom_path`	`str`	Relative path to DICOM file (e.g. `1/1_IM-0001-3001.dcm`)
`view_position`	`str`	`"frontal"`, `"lateral"`, or `"other"` (from figureId in XML)
`report`	`str`	Concatenated FINDINGS + IMPRESSION text
`findings`	`str`	FINDINGS section text
`impression`	`str`	IMPRESSION section text
`indication`	`str`	INDICATION section text
`mesh_major`	`str` (JSON)	Raw MeSH major terms as JSON list

Constructor arguments¶

Argument	Type	Required	Default	Description
`base_image_dir`	`str`	Yes*	`None`	Dataset root (contains `ecgen-radiology/` and `images/`)

Shared arguments (inherited from `BaseRadiologicalDataset`)¶

Argument	Type	Required	Default	Description
`output_cls`	`bool`	No	`False`	Include `"cls"` tensor in data dict
`output_mask`	`bool`	No	`False`	Include `"mask"` in data dict
`output_report`	`bool`	No	`False`	Include `"report"` string in data dict
`output_bbox`	`bool`	No	`False`	Include `"bbox"` and `"bbox_labels"` in data dict
`transform`	Compose	No	standard 2-D 224 px	MONAI Compose transform
`cache_dir`	`str`	No	`"./cache"`	MONAI cache directory. `None` disables
`dtype`	`torch.dtype`	No	`torch.bfloat16`	Output tensor dtype
`harmonized_df`	`pd.DataFrame`	No	`None`	Pre-built harmonized DataFrame
`harmonizer`	harmonizer	No	`None`	Pre-instantiated harmonizer
`harmonizer_path`	`str`	No	`None`	Path to saved harmonizer pickle

* Required unless harmonizer_path or harmonized_df is provided.

Dataset constructor¶

from radharmony.dataset import OpenICXRDataset

ds = OpenICXRDataset(
    base_image_dir="/data/OpenI-IU-CXR/",
    output_cls=True,
    output_report=True,
    cache_dir="./cache",
)

sample = ds.get_datasets()[0]
print(sample["img"].shape)   # torch.Size([3, 224, 224])  — RGB PNG
print(sample["cls"])         # tensor([0., 1., 0., ...])  — 8 binary labels
print(sample["report"])      # "Heart size is normal. ..."

Access DICOM paths via the harmonized DataFrame:

df = ds.get_harmonized_df()
print(df[["patient_id", "image_path", "dicom_path", "view_position"]].head(4))

Harmonizer¶

from radharmony.harmonizer import OpenICXRHarmonizer

h = OpenICXRHarmonizer(base_dir="/data/OpenI-IU-CXR/")
df = h.harmonize()
print(df.shape)           # (7470, 18)
print(df.columns.tolist())
# ['patient_id', 'study_id', 'image_path', 'view_position', 'report',
#  'no_finding', 'cardiomegaly', ..., 'dicom_path', 'findings', ...]

Harmonizer notes¶

No primary CSV — all data is parsed from ecgen-radiology/*.xml.
Images are RGB PNG — the img tensor has shape [3, 224, 224], not [1, 224, 224].
DICOM path derivation: XML parentImage id (e.g. CXR1_1_IM-0001-3001) maps to 1/1_IM-0001-3001.dcm using the regex CXR(\d+)_(?:\1_)?(.*) → {n}/{n}_{rest}.dcm. Two XML naming variants exist and are both handled.
MeSH label coverage: labels are only as complete as the MeSH annotations in the XML. Studies without MeSH tags get all zeros.