VinDr-PCXR Pediatric Chest X-ray¶

Modality: CXR | Format: DICOM | Dim: 2D | Labels: 15 binary conditions + bounding boxes

Overview¶

9,125 pediatric chest X-ray DICOMs (7,728 train + 1,397 test) from VinMec International Hospital, Vietnam. Annotated by experienced radiologists with image-level binary labels and bounding boxes for 37 thoracic conditions.

9,125 DICOMs — split into train/ (7,728) and test/ (1,397)
15 binary image-level labels — one annotation per image
~12,222 bounding boxes — pixel-space coords converted to normalized [y_min, y_max, x_min, x_max]

Download¶

Access requires credentialed PhysioNet account (free registration + DUA).

# Requires physionet.org account
wget -r -N -c -np --user <username> --ask-password \
  https://physionet.org/files/vindr-pcxr/1.0.0/ \
  -P ./VINDR-PCXR

Or via the PhysioNet web interface at https://physionet.org/content/vindr-pcxr/

Expected layout¶

VINDR-PCXR/                     ← base_image_dir points here
  train/
    6cb53aff85c71b98ad13d67a131708c6.dicom
    ...                          ← 7,728 DICOMs
  test/
    d7e71a052a753c3f2f3e317d60177bec.dicom
    ...                          ← 1,397 DICOMs
  image_labels_train.csv
  image_labels_test.csv
  annotations_train.csv
  annotations_test.csv

Label columns¶

15 binary conditions (image-level, one radiologist per image).

Column	Description
`no_finding`	No abnormality detected
`bronchitis`	Bronchitis
`brocho_pneumonia`	Broncho-pneumonia (as in original CSV)
`other_disease`	Other disease not listed
`bronchiolitis`	Bronchiolitis
`situs_inversus`	Situs inversus
`pneumonia`	Pneumonia
`pleuro_pneumonia`	Pleuro-pneumonia
`diagphramatic_hernia`	Diaphragmatic hernia (as in original CSV)
`tuberculosis`	Tuberculosis
`congenital_emphysema`	Congenital emphysema
`cpam`	Congenital pulmonary airway malformation
`hyaline_membrane_disease`	Hyaline membrane disease
`mediastinal_tumor`	Mediastinal tumor
`lung_tumor`	Lung tumor

Bounding box classes¶

37 classes from annotations_{split}.csv (stored as bbox_labels):

Anterior mediastinal mass, Aortic enlargement, Atelectasis, Boot-shaped heart, Bronchectasis, Bronchial thickening, Calcification, Cardiomegaly, Chest wall mass, Clavicle fracture, Consolidation, Dextro cardia, Diffuse aveolar opacity, Edema, Egg on string sign, Emphysema, Enlarged PA, Expanded edges of the anterior ribs, Infiltration, Interstitial lung disease - ILD, Intrathoracic digestive structure, Lung cavity, Lung cyst, Lung hyperinflation, Mediastinal shift, No finding, Other lesion, Other nodule/mass, Other opacity, Paraveterbral mass, Peribronchovascular interstitial opacity, Pleural effusion, Pleural thickening, Pneumothorax, Pulmonary fibrosis, Reticulonodular opacity, Stomach on the right side

Extra metadata columns¶

Column	Type	Description
`split`	`str`	`"train"` or `"test"`
`image_width`	`int`	DICOM image width in pixels
`image_height`	`int`	DICOM image height in pixels

Constructor arguments¶

Argument	Type	Required	Default	Description
`base_image_dir`	`str`	Yes*	`None`	Dataset root (contains `train/`, `test/`, CSV files)

Shared arguments (inherited from `BaseRadiologicalDataset`)¶

Argument	Type	Required	Default	Description
`output_cls`	`bool`	No	`False`	Include `"cls"` tensor (15 labels) in data dict
`output_mask`	`bool`	No	`False`	Not supported; ignored
`output_report`	`bool`	No	`False`	Not supported; ignored
`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 VinDrPCXRDataset

ds = VinDrPCXRDataset(
    base_image_dir="/data/VINDR-PCXR/",
    output_cls=True,
    output_bbox=True,
    cache_dir="./cache",
)

sample = ds.get_datasets()[0]
print(sample["img"].shape)        # torch.Size([1, 224, 224])
print(sample["cls"].shape)        # torch.Size([15])
print(sample["bbox"])             # [[y_min, y_max, x_min, x_max], ...]
print(sample["bbox_labels"])      # ["Bronchial thickening", ...]

Filter to train split only:

h = VinDrPCXRHarmonizer(base_dir="/data/VINDR-PCXR/")
df = h.harmonize()
train_df = df[df["split"] == "train"].reset_index(drop=True)

ds = VinDrPCXRDataset(
    base_image_dir="/data/VINDR-PCXR/",
    harmonized_df=train_df,
    output_cls=True,
)

Harmonizer¶

from radharmony.harmonizer import VinDrPCXRHarmonizer

h = VinDrPCXRHarmonizer(base_dir="/data/VINDR-PCXR/")
df = h.harmonize()
print(df.shape)           # (9125, 23)
print(df.columns.tolist())
# ['patient_id', 'study_id', 'image_path', 'bbox', 'bbox_labels',
#  'no_finding', 'bronchitis', ..., 'split', 'image_width', 'image_height']

Harmonizer notes¶

No PatientID/StudyUID in DICOM — patient_id and study_id are both set to image_id (the filename stem).
One annotation per image — each image was annotated by a single radiologist (identified by rad_ID).
Bbox normalization — DICOM dimensions are read from each image during harmonize() (header only, fast); bbox coords converted from pixel [x_min, y_min, x_max, y_max] to normalized [y_min, y_max, x_min, x_max].
All images have bboxes — all 9,125 images have at least one annotation in the annotations CSV.
Typos preserved — brocho_pneumonia and diagphramatic_hernia follow the original CSV column names verbatim.