EmoryCXR v2¶

Modality: CXR | Format: PNG | Dim: 2D | Labels: 14 pathologies | Access: Internal (Emory)

Overview¶

EmoryCXR v2 is an internal Emory de-identified chest radiography dataset containing ~2.4M images from ~360K patients across Emory Healthcare institutions. It includes:

Image-level metadata: patient demographics (age, sex, BMI, race, ethnicity), acquisition parameters (view position, frontal flag), institution, and study date.
Study-level finding labels: 14 CheXpert-aligned pathology labels (binary 0/1/NaN; no uncertain -1).
Study-level de-identified reports: free-text radiology reports stored inline in a CSV (not as separate files).

Labels and reports are stored at the accession (study) level and are automatically broadcast to all images within each study by the harmonizer.

Label columns¶

14 CheXpert-aligned findings (binary 0/1/NaN, no -1 uncertain):

atelectasis, cardiomegaly, consolidation, edema, enlarged_cardiomediastinum, fracture, lung_lesion, lung_opacity, no_finding, pleural_effusion, pleural_other, pneumonia, pneumothorax, support_devices

Extra metadata columns¶

The following demographic and acquisition columns are passed through to the harmonized DataFrame:

Column	Description
`sex`	Patient sex
`age`	Patient age
`bmi`	Body mass index
`race`	Patient race
`ethnicity`	Patient ethnicity
`frontal`	Whether the view is frontal (bool)
`institution`	De-identified institution name
`study_date`	De-identified study date

Constructor arguments¶

Argument	Type	Required	Default	Description
`base_image_dir`	`str`	Yes*	`None`	PNG root (contains patient sub-folders)
`csv_path`	`str`	No	auto	Path to metadata CSV
`label_csv_path`	`str`	No	`None`	Path to finding label CSV; required for `output_cls=True`
`report_csv_path`	`str`	No	`None`	Path to report CSV; required for `output_report=True`

Shared arguments (inherited from `BaseRadiologicalDataset`)¶

Argument	Type	Required	Default	Description
`output_cls`	`bool`	No	`False`	Yield 14-class label tensor under `cls`
`output_mask`	`bool`	No	`False`	Not supported; silently ignored
`output_report`	`bool`	No	`False`	Yield de-identified report text under `report`
`output_bbox`	`bool`	No	`False`	Not supported; silently ignored
`transform`	`Compose`	No	2D default	MONAI transform pipeline
`cache_dir`	`str`	No	`./cache`	PersistentDataset cache root; `None` disables caching
`dtype`	`torch.dtype`	No	`torch.bfloat16`	Image 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 harmonized CSV

*Required unless harmonizer_path, harmonized_df, or harmonizer is provided.

Dataset constructor¶

from radharmony.dataset import EmoryCXRDataset

ds = EmoryCXRDataset(
    base_image_dir="/path/to/EmoryCXRv2/DEID_PNG",
    csv_path="/path/to/EmoryCXRv2/TABLES/metadata.csv",
    label_csv_path="/path/to/EmoryCXRv2/TABLES/finding_labels.csv",
    report_csv_path="/path/to/EmoryCXRv2/TABLES/reports.csv",
    output_cls=True,
    output_report=True,
    cache_dir="./cache",
)
sample = ds[0]
# sample["img"]    → image tensor (C, H, W)
# sample["cls"]    → 14-class label tensor
# sample["report"] → de-identified report string

Harmonizer¶

from radharmony.harmonizer import EmoryCXRHarmonizer

h = EmoryCXRHarmonizer(
    csv_path="/path/to/EmoryCXRv2/TABLES/metadata.csv",
    base_image_dir="/path/to/EmoryCXRv2/DEID_PNG",
    label_csv_path="/path/to/EmoryCXRv2/TABLES/finding_labels.csv",
    report_csv_path="/path/to/EmoryCXRv2/TABLES/reports.csv",
)
df = h.harmonize()
# df columns: patient_id, study_id, image_path,
#             view_position, report,
#             atelectasis, cardiomegaly, ..., support_devices,
#             sex, age, bmi, race, ethnicity, frontal, institution, study_date

Harmonizer notes¶

Labels are at study level; they are broadcast to all SOP images in the same accession via a left merge. Images whose accession has no label entry will have NaN for all label columns.
Reports are stored as inline text in the report CSV (not as file paths). The report column in the harmonized DataFrame contains the de-identified text directly.
Masks and bounding boxes are not available in this release.

Example paths¶

Resource	Path
Images (PNG)	`/path/to/EmoryCXRv2/DEID_PNG/`
Metadata CSV	`/path/to/EmoryCXRv2/TABLES/metadata.csv`
Label CSV	`/path/to/EmoryCXRv2/TABLES/finding_labels.csv`
Report CSV	`/path/to/EmoryCXRv2/TABLES/reports.csv`

Image paths follow the structure: <empi_anon>/<AccessionNumber_anon>/<SOP>.png relative to the PNG root.