VinDr-CXR¶

Modality: CXR | Format: DICOM | Dim: 2D | Labels: 28 findings

Overview¶

VinDr-CXR is a large-scale chest X-ray dataset from Vietnam containing 18,000 CXR scans annotated by a panel of experienced radiologists. It provides image-level classification labels and bounding box annotations for 28 critical findings. The dataset is split into train (15,000) and test (3,000) sets with separate annotation files.

Download¶

Available from PhysioNet: VinDr-CXR. Requires CITI training and signed DUA.

Expected layout¶

vindr-cxr/1.0.0/
  train/
    image_labels_train.csv
    annotations_train.csv
    <image_id>.dicom
  test/
    image_labels_test.csv
    annotations_test.csv
    <image_id>.dicom

Label columns¶

28 findings (same for train and test): aortic_enlargement, atelectasis, calcification, cardiomegaly, clavicle_fracture, consolidation, copd, edema, emphysema, enlarged_pa, ild, infiltration, lung_cavity, lung_cyst, lung_opacity, lung_tumor, mediastinal_shift, no_finding, nodule/mass, other_diseases, other_lesion, pleural_effusion, pleural_thickening, pneumonia, pneumothorax, pulmonary_fibrosis, rib_fracture, tuberculosis

Constructor arguments¶

Both VinDrCXRTrainDataset and VinDrCXRTestDataset share the same signature:

Argument	Type	Required	Default	Description
`base_image_dir`	`str`	Yes	`None`	Split-specific dir — `train/` or `test/` — flat directory of `<image_id>.dicom` files (e.g. `/data/VinDr-CXR/vindr-cxr/1.0.0/train/`)
`csv_path`	`str`	No	auto	Image-level labels CSV — `image_labels_train.csv` (train) / `image_labels_test.csv` (test); auto-discovered under `base_image_dir`, then sibling dirs
`bbox_csv_path`	`str`	No	auto	Bounding box annotations CSV; auto-discovered

Shared arguments (inherited from `BaseRadiologicalDataset`)¶

Argument	Type	Required	Default	Description
`output_cls`	`bool`	No	`False`	Include `"cls"` in data dict
`output_mask`	`bool`	No	`False`	Include `"mask"` in data dict
`output_report`	`bool`	No	`False`	Include `"report"` in data dict
`output_bbox`	`bool`	No	`False`	Include `"bbox"` and `"bbox_labels"` in data dict
`transform`	MONAI transform	No	`None`	MONAI Compose transform; `None` uses the default pipeline
`cache_dir`	`str`	No	`"./cache"`	MONAI cache directory
`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 harmonized CSV

Dataset constructor¶

Train split¶

import torch
from radharmony.dataset import VinDrCXRTrainDataset

ds = VinDrCXRTrainDataset(
    base_image_dir="/data/VinDr-CXR/vindr-cxr/1.0.0/train/",
    output_cls=True,
    output_bbox=True,
    dtype=torch.float32,
)
train_ds, val_ds = ds.get_datasets(n_splits=5)

Test split¶

from radharmony.dataset import VinDrCXRTestDataset

ds = VinDrCXRTestDataset(
    base_image_dir="/data/VinDr-CXR/vindr-cxr/1.0.0/test/",
    output_cls=True,
    output_bbox=True,
    dtype=torch.float32,
)

Harmonizer¶

from radharmony.harmonizer import VinDrCXRTrainHarmonizer

h = VinDrCXRTrainHarmonizer(
    csv_path="/data/VinDr-CXR/vindr-cxr/1.0.0/train/image_labels_train.csv",
    base_image_dir="/data/VinDr-CXR/vindr-cxr/1.0.0/train/",
    bbox_csv_path="/data/VinDr-CXR/vindr-cxr/1.0.0/train/annotations_train.csv",
)
df = h.harmonize()
print(df.columns.tolist())
df.to_csv("vindr_cxr_train_harmonized.csv", index=False)

Load from saved harmonized CSV¶

import pandas as pd
from radharmony.dataset import VinDrCXRTrainDataset

ds = VinDrCXRTrainDataset(
    base_image_dir="/data/VinDr-CXR/vindr-cxr/1.0.0/train/",
    harmonized_df=pd.read_csv("vindr_cxr_train_harmonized.csv"),
    output_cls=True,
)

Harmonizer notes¶

Image-level CSV provides multi-label classification; annotations CSV provides per-box labels
Boxes are in pixel space; harmonizer normalises to fractional coordinates
Multiple annotators may label the same image; labels are consolidated to consensus in the harmonizer

Outputs¶

Flag	Key	Shape	Notes
`output_cls=True`	`"cls"`	`(28,)`	Multi-label binary
`output_bbox=True`	`"bbox"`, `"bbox_labels"`	list	One box per finding instance

Example paths¶

Role	Path
`base_image_dir` (train)	`/path/to/VinDr-CXR/physionet.org/files/vindr-cxr/1.0.0/train/`
`base_image_dir` (test)	`/path/to/VinDr-CXR/physionet.org/files/vindr-cxr/1.0.0/test/`
`annotations_csv_path` (train)	`/path/to/VinDr-CXR/physionet.org/files/vindr-cxr/1.0.0/annotations/annotations_train.csv`
`annotations_csv_path` (test)	`/path/to/VinDr-CXR/physionet.org/files/vindr-cxr/1.0.0/annotations/annotations_test.csv`
image-level labels (train)	`/path/to/VinDr-CXR/physionet.org/files/vindr-cxr/1.0.0/annotations/image_labels_train.csv`
image-level labels (test)	`/path/to/VinDr-CXR/physionet.org/files/vindr-cxr/1.0.0/annotations/image_labels_test.csv`