Shenzhen Hospital CXR Set¶

Modality: CXR | Format: PNG | Dim: 2D | Labels: Tuberculosis (binary) + 19 TB finding types

Overview¶

The Shenzhen Hospital CXR Set is a tuberculosis screening dataset created by the National Library of Medicine (NLM/LHNCBC) in collaboration with Shenzhen No.3 People's Hospital, Guangdong Medical College, Shenzhen, China. Images were captured as part of daily out-patient clinical routine using a Philips DR Digital Diagnose system.

662 total images: 336 TB-positive, 326 normal
Resolution: approximately 3000 × 3000 px (varies per image)
Label encoding: embedded in each filename (_0 = normal, _1 = TB)
TB finding annotations: 19 binary finding-type columns + per-finding region masks (TB patients only)

This is the companion dataset to the Montgomery County CXR Set, from the same NLM TB screening collaboration.

Note

There is no primary CSV — ShenzhenCXRHarmonizer scans the CXR_png/ directory and builds the DataFrame from filenames.

Download¶

Available at NLM/LHNCBC Tuberculosis CXR Datasets. No authentication required.

The NLM server blocks directory listing on the root URL, so the dataset must be mirrored per-subdirectory:

BASE="https://data.lhncbc.nlm.nih.gov/public/Tuberculosis-Chest-X-ray-Datasets/Shenzhen-Hospital-CXR-Set"
DEST="./Shenzhen-Hospital-CXR-Set"

wget --mirror --no-parent --no-host-directories --cut-dirs=3 \
     --reject='index.html*' -e robots=off -P "$DEST" \
     "${BASE}/CXR_png/" "${BASE}/ClinicalReadings/"

wget -P "$DEST" \
     "${BASE}/NLM-ChinaCXRSet-ReadMe.docx" \
     "${BASE}/shenzhen_consensus_roi.csv"

For annotation files (Annotations/, Annotations-2/), index pages must be parsed to discover file names. See the Acquiring Datasets page for a complete download script.

Please cite the following publications if you use this dataset:

Jaeger S, et al. Automatic tuberculosis screening using chest radiographs. IEEE Trans Med Imaging. 2014;33(2):233–45.

Candemir S, et al. Lung segmentation in chest radiographs using anatomical atlases with nonrigid registration. IEEE Trans Med Imaging. 2014;33(2):577–90.

Expected layout¶

Shenzhen-Hospital-CXR-Set/
  CXR_png/                              ← base_image_dir points here
    CHNCXR_0001_0.png                   # Normal
    CHNCXR_0327_1.png                   # TB-positive
    ...                                 # 662 total
  ClinicalReadings/                     # sibling of CXR_png/
    CHNCXR_0001_0.txt                   # "male 45yrs\nnormal"
    CHNCXR_0327_1.txt                   # "female 28yrs\nSTB"
    ...
  Annotations-2/                        # sibling of CXR_png/
    Statistics_ShenzhenDataset_US.csv   # 19 binary TB finding columns (TB only)
    masks/
      CHNCXR_0330_1_Pleural_Effusion_2.png
      CHNCXR_0332_1_Linear_Density_1.png
      ...                               # ~4,000 per-finding region masks
  shenzhen_consensus_roi.csv
  NLM-ChinaCXRSet-ReadMe.docx

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 Annotations-2/ are read as sibling directories.

Label columns¶

Column	Type	Description
`tuberculosis`	`int` (0 or 1)	`0` = normal, `1` = active tuberculosis

TB finding annotation columns¶

Present only for TB-positive patients (0327–0662); NaN for normal patients. Values are counts of annotated regions (0 = absent, ≥1 = present).

Column	Prevalence (TB patients)
`Clustered_Nodule_(2mm-5mm_apart)`	605
`Linear_Density`	339
`Small_Infiltrate_(non-linear)`	219
`Single_Nodule_(non-calcified)`	193
`Calcified_Nodule`	171
`Pleural_Thickening_(non-apical)`	113
`Cavity`	101
`Moderate_Infiltrate_(non-linear)`	97
`Apical_Thickening`	92
`Pleural_Effusion`	36
`Adenopathy`	30
`Miliary`	27
`Retraction`	22
`Severe_Infiltrate_(Consolidation)`	20
`Other`	71
`Unknown`	19
`Thickening_of_interlobar_fissure`	8
`Calcified_lymph_node`	8
`Calcification_(other_than_nodule&lymphnod)`	6

Extra metadata columns¶

Column	Type	Description
`patient_age`	`int`	Patient age in years (from ClinicalReadings/)
`patient_sex`	`str`	`"M"` or `"F"` (from ClinicalReadings/)
`report`	`str`	Raw clinical reading text
`finding_masks`	`str` (JSON)	JSON list of relative paths (e.g. `"../Annotations-2/masks/..."`) to per-finding region masks; `None` for normal patients

Constructor arguments¶

Argument	Type	Required	Default	Description
`base_image_dir`	`str`	Yes*	`None`	The `CXR_png/` images directory itself
`mask_output_dir`	`str`	No	`None`	Required when `output_mask=True` — directory the unioned TB-region PNGs are written to
`mask_num_cores`	`int`	No	`1`	Reserved for signature parity with SIIM-ACR PTX; currently unused

Shared arguments (inherited from `BaseRadiologicalDataset`)¶

Argument	Type	Required	Default	Description
`output_cls`	`bool`	No	`False`	Yield TB label under `"cls"`
`output_mask`	`bool`	No	`False`	Yield fused per-finding "TB region" mask under `"mask"`; requires `mask_output_dir`
`output_report`	`bool`	No	`False`	Yield clinical reading text under `"report"`
`output_bbox`	`bool`	No	`False`	Not supported; ignored
`transform`	Compose	No	standard 2-D 224 px	MONAI Compose transform; `output_keys` extended with `mask` when `output_mask=True`
`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 a saved harmonized pickle

* Required unless harmonizer_path or harmonized_df is provided.

Mask output¶

When output_mask=True, the dataset class invokes ShenzhenCXRHarmonizer.preprocess_masks(mask_output_dir, ...) which builds one binary "TB region" PNG per image:

TB-positive cases: union the per-finding masks listed in finding_masks (fused = OR(mask_i for i in findings)).
TB-negative cases: write an all-zero PNG of the same size as the source image (matches the SIIM-ACR PTX convention so the row stays in the stream).

The mask_path column is populated with the absolute path to each fused PNG. The loop is idempotent — existing PNGs are not rewritten.

Passing output_mask=True without mask_output_dir raises ValueError.

Dataset constructor¶

from radharmony.dataset import ShenzhenCXRDataset

ds = ShenzhenCXRDataset(
    base_image_dir="/data/Shenzhen-Hospital-CXR-Set/CXR_png/",
    output_cls=True,
    output_report=True,
    cache_dir="./cache",
)

sample = ds[0]
print(sample["img"].shape)   # torch.Size([1, 224, 224])
print(sample["cls"])         # tensor([0.]) or tensor([1.])

# With segmentation
ds_seg = ShenzhenCXRDataset(
    base_image_dir="/data/Shenzhen-Hospital-CXR-Set/CXR_png/",
    output_cls=True,
    output_mask=True,
    mask_output_dir="/scratch/shenzhen_tb_region_masks/",
)

Access TB finding annotations and mask paths via the harmonized DataFrame:

df = ds.harmonized_df
tb = df[df.tuberculosis == 1]
print(tb[["patient_id", "Cavity", "Pleural_Effusion", "finding_masks"]].head())

import json
masks = json.loads(tb.iloc[0]["finding_masks"])
# ['../Annotations-2/masks/CHNCXR_0327_1_Small_Infiltrate_(non-linear)_1.png', ...]

Harmonizer¶

from radharmony.harmonizer import ShenzhenCXRHarmonizer

h = ShenzhenCXRHarmonizer(base_dir="/data/Shenzhen-Hospital-CXR-Set/CXR_png/")
df = h.harmonize()
print(df.shape)                           # (662, 27)
print(df["tuberculosis"].value_counts())  # 1: 336, 0: 326

Harmonizer notes¶

No primary CSV: labels are derived from filename suffix (_0 / _1).
Clinical readings: ClinicalReadings/<stem>.txt format — line 1 is "<sex> <age>yrs", remaining lines are diagnosis text.
TB findings: joined from Annotations-2/Statistics_ShenzhenDataset_US.csv; NaN for normal patients.
Finding masks: ~4,000 per-finding region mask PNGs in Annotations-2/masks/; stored as a JSON list per patient in finding_masks (paths are sibling-relative, i.e. start with "../Annotations-2/...").
TB-negative sample retention: with output_mask=True, normals get a black-PNG mask rather than being dropped (matches SIIM-ACR PTX convention).
Companion dataset: shares filename convention with Montgomery County CXR.