Emory CHORUS (X-ray subset)

Modality: CXR | Format: DICOM | Dim: 2D | Labels: none (image-only) | Access: Internal (Emory)

Overview

Emory CHORUS is an internal Emory dataset that pairs an OMOP-CDM clinical warehouse (patients, conditions, measurements, procedures, visits) with a multi-modality DICOM imaging tree. RadHarmony serves the curated X-ray subset and, by default, filters it to chest radiographs.

This release is image-only: no finding labels, masks, bounding boxes, or reports ship with it. (Labels could be derived from the OMOP condition codes in a future step; that is out of scope here.) patient_id (DICOM PatientID) groups a patient's studies for patient-level cross-validation.

Unlike every other RadHarmony dataset, CHORUS ships no metadata CSV — the study/series/image table is built by walking the DICOM tree and reading each file's header. That scan is expensive, so the harmonizer caches the result to a manifest CSV: the first run scans and writes it; later runs read it back.

The X-ray subset spans ~1,500 patients. Filtered to chest radiographs it is about 10.3K images / 4.9K studies / 1.4K patients (all DX, MONOCHROME2); the remainder are abdomen, spine, skull, and extremity films, kept only when chest_only=False.

Extra metadata columns

Carried through to the harmonized DataFrame so studies can be filtered or stratified without re-scanning the tree:

Column	Description
series_id	DICOM SeriesInstanceUID
view_position	DICOM ViewPosition (AP, PA, LL, LATERAL, …)
body_part	DICOM BodyPartExamined (e.g. CHEST, PORT CHEST)
study_date	DICOM StudyDate
sex	DICOM PatientSex
age	DICOM PatientAge
manufacturer	DICOM Manufacturer

Constructor arguments

Argument	Type	Required	Default	Description
base_image_dir	str	Yes*	None	DICOM root (images_dicom_xray, contains per-patient folders)
manifest_csv_path	str	No	None	Manifest scan cache; read if present, else scanned and written here
chest_only	bool	No	True	Keep only chest radiographs (DX/CR with CHEST in body-part/description)
num_workers	int	No	12	Parallel worker processes for the header scan

Shared arguments (inherited from BaseRadiologicalDataset)

Argument	Type	Required	Default	Description
output_cls	bool	No	False	Not supported (no labels); silently ignored
output_mask	bool	No	False	Not supported; silently ignored
output_report	bool	No	False	Not supported; silently ignored
output_bbox	bool	No	False	Not supported; silently ignored
transform	Compose	No	2D default	MONAI transform pipeline (DICOM VOI-LUT / MONOCHROME handled)
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 EmoryCHORUSDataset

ds = EmoryCHORUSDataset(
    base_image_dir="/path/to/Emory_CHORUS/images_dicom_xray",
    manifest_csv_path="/path/to/Emory_CHORUS/chorus_xray_manifest.csv",
    cache_dir="./cache",
)
sample = ds[0]
# sample["img"] → image tensor (C, H, W); image-only, no other keys

Harmonizer

from radharmony.harmonizer import EmoryCHORUSHarmonizer

h = EmoryCHORUSHarmonizer(
    base_image_dir="/path/to/Emory_CHORUS/images_dicom_xray",
    csv_path="/path/to/Emory_CHORUS/chorus_xray_manifest.csv",  # scan cache
    chest_only=True,
)
df = h.harmonize()
# df columns: patient_id, study_id, image_path,
#             series_id, view_position,
#             body_part, study_date, sex, age, manufacturer

Harmonizer notes

• No metadata CSV. The harmonizer scans DICOM headers to build the study / series / image table. Pass manifest_csv_path (dataset) / csv_path (harmonizer) so the scan is cached and reused; without it the tree is re-scanned on every construction.
• Chest filter. chest_only=True keeps Modality in {DX, CR} with CHEST in the body-part or study description. This admits combined chest+abdomen and rib studies whose description mentions the chest.
• Image-only. No labels, masks, bounding boxes, or reports in this release, so only the img key is produced.

Example paths

Resource	Path
Images (DICOM)	/path/to/Emory_CHORUS/images_dicom_xray/
Manifest CSV (scan cache)	user-chosen; written on first harmonize
OMOP-CDM tables	/path/to/Emory_CHORUS/tables/ (not used by this release)

Image paths follow the structure <person_id>/Images/<StudyInstanceUID>/<SeriesInstanceUID>/<sop>.dcm relative to the image root.