SIIM-ACR Pneumothorax¶

Modality: CXR | Format: DICOM | Dim: 2D | Labels: pneumothorax (binary) + mask

Overview¶

The SIIM-ACR Pneumothorax Segmentation dataset (2019 Kaggle challenge) contains ~12,047 training chest X-rays; the train-rle.csv has 12,954 RLE rows because some images carry multiple pneumothorax annotations. Each image is labelled for the presence of pneumothorax, and positive cases include one or more pixel-level segmentation masks encoded as run-length encoded (RLE) strings in the CSV.

Download¶

Available on Kaggle: SIIM-ACR Pneumothorax Segmentation. Requires Kaggle account.

The harmonizer globs **/*.dcm under dicom_dir and keeps the last three path components (<study>/<series>/<image>.dcm) as image_path.

Expected layout¶

SIIM_ACR_Pneumothorax/
  train-rle.csv
  dicom-images-train/                    # ← base_image_dir for the train split
    <StudyInstanceUID>/
      <SeriesInstanceUID>/
        <image_id>.dcm
  dicom-images-test/                     # ← base_image_dir for the test split
    <StudyInstanceUID>/
      <SeriesInstanceUID>/
        <image_id>.dcm

Label columns¶

Column	Description
`pneumothorax`	1 = pneumothorax present, 0 = absent

Constructor arguments¶

Argument	Type	Required	Default	Description
`base_image_dir`	`str`	Yes	`None`	`dicom-images-train/` (or `dicom-images-test/` for the test split) — root of the DICOM tree (passed internally to the harmonizer as `dicom_dir`)
`csv_path`	`str`	No	auto	`train-rle.csv`; auto-discovered
`mask_output_dir`	`str`	No	`None`	Directory to save decoded PNG masks; required for `output_mask=True`
`mask_num_cores`	`int`	No	`1`	Parallel workers for RLE mask decoding

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¶

The constructor arguments above apply to both SIIMACRPTXTrainDataset and SIIMACRPTXTestDataset. The test split has no labels — output_cls, output_mask, and output_bbox are silently ignored.

Train split¶

import torch
from radharmony.dataset import SIIMACRPTXTrainDataset

ds = SIIMACRPTXTrainDataset(
    base_image_dir="/data/SIIM_ACR_Pneumothorax/dicom-images-train/",
    mask_output_dir="/data/SIIM_ACR_Pneumothorax/masks/",
    output_cls=True,
    output_mask=True,
    dtype=torch.float32,
)
train_ds, val_ds = ds.get_datasets(n_splits=5)

Test split¶

from radharmony.dataset import SIIMACRPTXTestDataset

ds = SIIMACRPTXTestDataset(
    base_image_dir="/data/SIIM_ACR_Pneumothorax/dicom-images-test/",
    dtype=torch.float32,
)

Harmonizer¶

from radharmony.harmonizer import SIIMACRPTXTrainHarmonizer

h = SIIMACRPTXTrainHarmonizer(
    csv_path="/data/SIIM_ACR_Pneumothorax/train-rle.csv",
    dicom_dir="/data/SIIM_ACR_Pneumothorax/dicom-images-train/",
)
df = h.harmonize()
print(df.columns.tolist())
df.to_csv("siim_ptx_harmonized.csv", index=False)

Load from saved harmonized CSV¶

import pandas as pd
from radharmony.dataset import SIIMACRPTXTrainDataset

ds = SIIMACRPTXTrainDataset(
    base_image_dir="/data/SIIM_ACR_Pneumothorax/dicom-images-train/",
    harmonized_df=pd.read_csv("siim_ptx_harmonized.csv"),
    output_cls=True,
)

Harmonizer notes¶

RLE masks in train-rle.csv use the value -1 to indicate no pneumothorax (no mask)
When output_mask=True, the harmonizer decodes RLE strings to PNG files in mask_output_dir on first run
Multiple RLE strings per image (multiple opacities) are merged into a single binary mask

Outputs¶

Flag	Key	Shape	Notes
`output_cls=True`	`"cls"`	`(1,)`	Binary: pneumothorax present/absent
`output_mask=True`	`"mask"`	`(1, H, W)`	Decoded binary segmentation mask

Example paths¶

Role	Path
`base_image_dir` (train)	`/path/to/SIIM_ACR_Pneumothorax/dicom-images-train/`
`base_image_dir` (test)	`/path/to/SIIM_ACR_Pneumothorax/dicom-images-test/`
`csv_path` (`train-rle.csv`)	`/path/to/SIIM_ACR_Pneumothorax/train-rle.csv`
pre-decoded mask dir	`/path/to/SIIM_ACR_Pneumothorax/masks`