Montgomery County CXR (TB)¶

Modality: CXR | Format: PNG | Dim: 2D | Labels: TB classification (binary) + lung segmentation masks

Overview¶

The Montgomery County Tuberculosis Chest X-ray dataset was collected in collaboration with the National Library of Medicine (LHNCBC) and the Montgomery County Department of Health and Human Services (Maryland, USA). It contains 138 posterior-anterior (PA) chest radiographs — 80 normal and 58 TB-positive — with binary TB labels, patient demographics, and manually drawn left/right lung segmentation masks.

Despite its small size, the dataset is widely used for TB classification and lung segmentation benchmarking.

Download¶

Available from the LHNCBC Tuberculosis Chest X-ray Datasets collection:

https://data.lhncbc.nlm.nih.gov/public/Tuberculosis-Chest-X-ray-Datasets/Montgomery-County-CXR-Set/

base_image_dir points at the CXR_png/ images directory itself — the last common folder containing every image (matching the convention used by other RadHarmony datasets). ClinicalReadings/ and ManualMask/ are read as sibling directories.

Downloading¶

The dataset is publicly available — no account required. Download and unzip with:

BASE_URL="https://data.lhncbc.nlm.nih.gov/public/Tuberculosis-Chest-X-ray-Datasets/Montgomery-County-CXR-Set"

wget -q "${BASE_URL}/MontgomerySet.zip" -O MontgomerySet.zip
unzip -q MontgomerySet.zip
# → produces MontgomerySet/ in the current directory

If the zip URL redirects or is unavailable, visit the index page and follow the download link:

https://data.lhncbc.nlm.nih.gov/public/Tuberculosis-Chest-X-ray-Datasets/Montgomery-County-CXR-Set/index.html

Expected layout¶

MontgomerySet/
  CXR_png/                   # base_image_dir points here
    MCUCXR_0001_0.png        # normal  (suffix _0)
    MCUCXR_0001_1.png        # TB      (suffix _1)
    ...
  ManualMask/                # sibling of CXR_png/
    leftMask/
      MCUCXR_0001_0.png
      ...
    rightMask/
      MCUCXR_0001_0.png
      ...
  ClinicalReadings/          # sibling of CXR_png/
    MCUCXR_0001_0.txt        # one file per patient
    ...

Label columns¶

Column	Type	Description
`tuberculosis`	int (0/1)	Tuberculosis status — 0 = normal, 1 = TB-positive

Extra metadata columns¶

Column	Description
`sex`	Patient sex (M/F)
`age`	Patient age in years
`mask_path_left`	Relative path to left lung mask PNG (if present)
`mask_path_right`	Relative path to right lung mask PNG (if present)

Constructor arguments (harmonizer)¶

Argument	Type	Required	Default	Description
`base_dir`	`str`	Yes	—	The `CXR_png/` images directory itself (sibling `ClinicalReadings/` and `ManualMask/` are read relative to it)

MontgomeryCXRHarmonizer.harmonize() also accepts mask_output_dir and mask_num_cores kwargs (see Mask output).

Constructor arguments (dataset)¶

Argument	Type	Default	Description
`base_image_dir`	`str`	`None`	The `CXR_png/` images directory
`mask_output_dir`	`str`	`None`	Required when `output_mask=True` — directory the fused left+right lung-mask PNGs are written to
`mask_num_cores`	`int`	`1`	Reserved for signature parity with SIIM-ACR PTX; currently unused (fusion loop is small)

Shared arguments (inherited from `BaseRadiologicalDataset`)¶

Argument	Type	Default	Description
`output_cls`	`bool`	`False`	Yield TB label under the `"cls"` key
`output_mask`	`bool`	`False`	Yield fused lung mask under the `"mask"` key; requires `mask_output_dir`
`output_report`	`bool`	`False`	Yield clinical reading text under the `"report"` key
`output_bbox`	`bool`	`False`	Not supported; ignored
`transform`	MONAI Compose	standard 224 px 2-D	Default 2-D pipeline; `output_keys` extended with `mask` when `output_mask=True`
`cache_dir`	`str`	`"./cache"`	MONAI `PersistentDataset` cache; `None` disables
`dtype`	`torch.dtype`	`torch.bfloat16`	Output tensor dtype
`harmonized_df`	`pd.DataFrame`	`None`	Pre-built harmonized DataFrame
`harmonizer`	harmonizer	`None`	Pre-instantiated harmonizer
`harmonizer_path`	`str`	`None`	Path to saved harmonized CSV

Mask output¶

When output_mask=True, the dataset class invokes MontgomeryCXRHarmonizer.preprocess_masks(mask_output_dir, ...) which fuses the left and right manual lung masks into a single binary PNG per image (fused = (left > 0) | (right > 0)) and writes one file per image into mask_output_dir. The harmonized DataFrame's mask_path column is then populated with the absolute path to each fused PNG, so MONAI can load it directly. The fusion loop is idempotent — existing fused PNGs are not rewritten on subsequent runs.

Passing output_mask=True without mask_output_dir raises ValueError.

Dataset constructor¶

from radharmony.harmonizer import MontgomeryCXRHarmonizer
from radharmony.dataset import MontgomeryCXRDataset

# Harmonizer-only flow
h = MontgomeryCXRHarmonizer(base_dir="/data/MontgomerySet/CXR_png/")
df = h.harmonize()
print(df.shape)                       # (138, ...)
print(df["tuberculosis"].value_counts())   # 0: 80, 1: 58
print(df.columns.tolist())

# Dataset flow with classification + lung mask output
ds = MontgomeryCXRDataset(
    base_image_dir="/data/MontgomerySet/CXR_png/",
    output_cls=True,
    output_mask=True,
    mask_output_dir="/scratch/montgomery_fused_masks/",
)

Harmonizer notes¶

The filename suffix (_0 / _1) encodes TB status; this is the authoritative source of the tuberculosis label.
Lung masks are optional — mask_path_left and mask_path_right are NaN if ManualMask/ is absent. In that case output_mask=True rows will be dropped from the data stream.
ClinicalReadings/<stem>.txt is one file per patient (free-text plus Patient's Sex / Patient's Age headers); the harmonizer parses sex, age, and report from it.