Skip to content

RadHarmony

RSNA Pneumonia (Kaggle)

f10409/RadHarmony

RSNA Pneumonia Detection (Kaggle Stage 2)¶

Modality: CXR | Format: DICOM | Dim: 2D | Labels: 3 classes + bboxes

Overview¶

The RSNA Pneumonia Detection Challenge Kaggle Stage 2 dataset contains 26,684 chest X-rays distributed as two flat CSV files and a flat DICOM directory. Same underlying images as RSNA Pneumonia, but uses the Kaggle CSV distribution format instead of the adjudicated JSON.

Download¶

Available on Kaggle: RSNA Pneumonia Detection Challenge. Requires Kaggle account.

Expected layout¶

rsna-pneumonia-detection-challenge/
  stage_2_detailed_class_info.csv
  stage_2_train_labels.csv
  stage_2_train_images/
    <patientId>.dcm

Label columns¶

Column	Description
`lung_opacity`	Lung opacity (pneumonia) present
`no_lung_opacity_/_not_normal`	Abnormal but no lung opacity
`normal`	Normal chest X-ray

Constructor arguments¶

Argument	Type	Required	Default	Description
`base_image_dir`	`str`	Yes	`None`	`stage_2_train_images/` (or `stage_2_test_images/` for the test split) — flat dir of `<patientId>.dcm` files
`csv_path`	`str`	No	auto	`stage_2_detailed_class_info.csv`; auto-discovered
`bbox_csv_path`	`str`	No	auto	`stage_2_train_labels.csv`; required for bounding boxes

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¶

import torch
from radharmony.dataset import (
    RSNAPneumoniaKaggleTrainDataset,
    RSNAPneumoniaKaggleTestDataset,
)

# Train split (with class labels and bboxes)
ds_train = RSNAPneumoniaKaggleTrainDataset(
    base_image_dir="/data/rsna-pneumonia-detection-challenge/stage_2_train_images/",
    output_cls=True,
    output_bbox=True,
    dtype=torch.float32,
)
train_ds, val_ds = ds_train.get_datasets(n_splits=5)

# Test split (images only — no labels)
ds_test = RSNAPneumoniaKaggleTestDataset(
    base_image_dir="/data/rsna-pneumonia-detection-challenge/stage_2_test_images/",
    dtype=torch.float32,
)

Harmonizer¶

from radharmony.harmonizer import RSNAPneumoniaKaggleHarmonizer

h = RSNAPneumoniaKaggleHarmonizer(
    csv_path="/data/rsna-pneumonia-detection-challenge/stage_2_detailed_class_info.csv",
    base_image_dir="/data/rsna-pneumonia-detection-challenge/stage_2_train_images/",
    bbox_csv_path="/data/rsna-pneumonia-detection-challenge/stage_2_train_labels.csv",
)
df = h.harmonize()
print(df.columns.tolist())
# ['patient_id', 'study_id', 'image_path', 'lung_opacity',
#  'no_lung_opacity_/_not_normal', 'normal', 'bbox', 'bbox_labels']
df.to_csv("rsna_pneumonia_kaggle_harmonized.csv", index=False)

Load from saved harmonized CSV¶

import pandas as pd
from radharmony.dataset import RSNAPneumoniaKaggleTrainDataset

ds = RSNAPneumoniaKaggleTrainDataset(
    base_image_dir="/data/rsna-pneumonia-detection-challenge/stage_2_train_images/",
    harmonized_df=pd.read_csv("rsna_pneumonia_kaggle_harmonized.csv"),
    output_cls=True,
)

Harmonizer notes¶

stage_2_detailed_class_info.csv provides class labels (class column: "Normal", "No Lung Opacity / Not Normal", "Lung Opacity")
stage_2_train_labels.csv provides bounding boxes for Target=1 rows; Target=0 rows have NaN coordinates
All images are 1024×1024 DICOM; bounding boxes are normalised by dividing by 1024
Bboxes aggregated per patient: images with multiple opacities have a list of boxes
Normal / No Lung Opacity images get bbox=[] and bbox_labels=[]

Outputs¶

Flag	Key	Shape	Notes
`output_cls=True`	`"cls"`	`(3,)`	One-hot: lung_opacity, no_lung_opacity, normal
`output_bbox=True`	`"bbox"`, `"bbox_labels"`	list	Empty list for non-opacity cases

Example paths¶

Role	Path
`base_image_dir` (train)	`/path/to/rsna-pneumonia-detection-challenge/stage_2_train_images/`
`base_image_dir` (test)	`/path/to/rsna-pneumonia-detection-challenge/stage_2_test_images/`
`csv_path` (`stage_2_train_labels.csv`)	`/path/to/rsna-pneumonia-detection-challenge/stage_2_train_labels.csv`