BRAX¶

Modality: CXR | Format: DICOM + PNG | Dim: 2D | Labels: 14 pathologies

Overview¶

BRAX (Brazilian Chest X-Ray) v1.1.0 is a PhysioNet credentialed-access release containing ~40,967 chest X-ray images from 19,351 patients across a Brazilian hospital network. Labels are CheXpert-aligned (14 classes), NLP-derived from Portuguese radiology reports using the original CheXpert labeller.

Both DICOM and PNG variants ship in the same download. RadHarmony provides two dataset classes sharing one harmonizer:

BRAXDataset — DICOM images under Anonymized_DICOMs/
BRAXPNGDataset — PNG images under images/ (faster to load for large training runs)

Download¶

Available at PhysioNet: BRAX. Requires credentialed access and a signed DUA.

base_image_dir should point at the BRAX root (the directory containing master_spreadsheet.csv and both image subdirectories).

Expected layout¶

brax/1.1.0/
  master_spreadsheet.csv          # or master_spreadsheet_update.csv
  Anonymized_DICOMs/
    id_<PatientID>/Study_<UID>/Series_<UID>/image-<UID>.dcm
  images/
    id_<PatientID>/Study_<UID>/Series_<UID>/image-<UID>.png

Label columns¶

14 CheXpert-aligned findings: atelectasis, cardiomegaly, consolidation, edema, enlarged_cardiomediastinum, fracture, lung_lesion, lung_opacity, no_finding, pleural_effusion, pleural_other, pneumonia, pneumothorax, support_devices

Label cells use the encoding: 1 = positive, 0 = negation, -1 = uncertain, NaN = not mentioned.

Uncertain label strategies¶

The uncertain_strategy parameter controls how -1 (uncertain) cells are handled:

Strategy	`-1` →	`NaN` →	Rows dropped
`raw` (default)	preserved	preserved	none
`u_zeros`	`0`	`0`	none
`u_ones`	`1`	`0`	none
`u_ignore`	`NaN`	`0`	none
`drop`	`NaN`	`0`	rows with any NaN label

Use u_zeros for standard training-ready 0/1 tensors. Use raw when you need to faithfully reproduce the source encoding.

Note

Reports are not distributed in BRAX 1.1.0. output_report=True is ignored with a warning.

Constructor arguments¶

Argument	Type	Required	Default	Description
`base_image_dir`	`str`	Yes	`None`	BRAX root (contains `master_spreadsheet.csv`, `Anonymized_DICOMs/`, `images/`)
`csv_path`	`str`	No	auto	Path to `master_spreadsheet.csv`; auto-discovered
`uncertain_strategy`	`str`	No	`"raw"`	One of `raw`, `u_zeros`, `u_ones`, `u_ignore`, `drop`

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¶

DICOM variant¶

import torch
from radharmony.dataset import BRAXDataset

ds = BRAXDataset(
    base_image_dir="/data/physionet.org/files/brax/1.1.0/",
    uncertain_strategy="u_zeros",
    output_cls=True,
    dtype=torch.float32,
)
train_ds, val_ds = ds.get_datasets(n_splits=10)

PNG variant¶

from radharmony.dataset import BRAXPNGDataset

ds = BRAXPNGDataset(
    base_image_dir="/data/physionet.org/files/brax/1.1.0/",
    uncertain_strategy="u_zeros",
    output_cls=True,
)
train_ds, val_ds = ds.get_datasets(n_splits=10)

Harmonizer¶

from radharmony.harmonizer import BRAXHarmonizer

h = BRAXHarmonizer(
    csv_path="/data/physionet.org/files/brax/1.1.0/master_spreadsheet.csv",
    base_image_dir="/data/physionet.org/files/brax/1.1.0/",
    image_format="dicom",       # or "png"
    uncertain_strategy="u_zeros",
)
df = h.harmonize()
print(df.columns.tolist())
df.to_csv("brax_harmonized.csv", index=False)

Load from saved harmonized CSV¶

import pandas as pd
from radharmony.dataset import BRAXDataset

ds = BRAXDataset(
    base_image_dir="/data/physionet.org/files/brax/1.1.0/",
    harmonized_df=pd.read_csv("brax_harmonized.csv"),
    output_cls=True,
)

Outputs¶

Flag	Key	Shape	Notes
`output_cls=True`	`"cls"`	`(14,)`	CheXpert-aligned labels

Harmonizer notes¶

CSV auto-discovery checks for master_spreadsheet.csv then master_spreadsheet_update.csv
image_path in the harmonized DataFrame is relative to base_image_dir and includes the Anonymized_DICOMs/ or images/ prefix as written in the source CSV
series_id is extracted from the Series_<UID> segment of the image path via regex
Extra demographic columns (patient_sex, patient_age, manufacturer, study_date) flow through as metadata; patient_age is binned into 5-year groups in the de-identified release (treat as ordinal, not continuous)
Both frontal (AP/PA) and lateral views are retained; filter via the harmonized DataFrame if frontal-only is needed