RANZCR CLiP¶

Modality: CXR | Format: JPEG | Dim: 2D | Labels: 11 catheter/line position labels

Overview¶

RANZCR CLiP (Royal Australian and New Zealand College of Radiologists Catheter and Line in Position) is a Kaggle competition dataset containing ~30,000 frontal chest X-rays annotated for the presence and placement of central venous catheters (CVC), endotracheal tubes (ETT), nasogastric tubes (NGT), and Swan-Ganz catheters. The primary task is multi-label binary classification.

A subset of training images also includes polyline annotations tracing each catheter/line, which RadHarmony can rasterize into binary masks.

Download¶

Available via Kaggle: ranzcr-clip-catheter-line-classification. Accept the competition rules before downloading.

import kagglehub
path = kagglehub.competition_download("ranzcr-clip-catheter-line-classification")

Expected layout¶

ranzcr-clip-catheter-line-classification/
  train.csv
  train_annotations.csv
  sample_submission.csv
  train/
    <StudyInstanceUID>.jpg    # ~30k frontal CXRs with labels
  test/
    <StudyInstanceUID>.jpg    # competition holdout, no public labels

Label columns¶

11 binary multi-label targets:

Label	Description
`cvc_abnormal`	CVC in abnormal position
`cvc_borderline`	CVC in borderline position
`cvc_normal`	CVC in normal position
`ett_abnormal`	ETT in abnormal position
`ett_borderline`	ETT in borderline position
`ett_normal`	ETT in normal position
`ngt_abnormal`	NGT in abnormal position
`ngt_borderline`	NGT in borderline position
`ngt_incompletely_imaged`	NGT incompletely imaged
`ngt_normal`	NGT in normal position
`swan_ganz_catheter_present`	Swan-Ganz catheter present

Test split rows have NaN for all label columns (no public labels released).

Constructor arguments¶

Argument	Type	Required	Default	Description
`base_image_dir`	`str`	Yes	`None`	Kaggle download root (contains `train/`, `test/`, and CSVs)
`csv_path`	`str`	No	auto	Path to `train.csv`; auto-discovered
`annotations_csv_path`	`str`	No	auto	Path to `train_annotations.csv`; required only for `output_mask=True`
`mask_output_dir`	`str`	No	`None`	Directory for rasterized polyline PNGs; required for `output_mask=True`
`mask_line_thickness`	`int`	No	`15`	Pixel width for `cv2.polylines` when rasterizing
`include_test_split`	`bool`	No	`True`	Include test rows (NaN labels) in harmonized DataFrame
`mask_num_cores`	`int`	No	`1`	Worker threads for mask pre-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 rasterized polyline mask under `"mask"`
`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¶

Classification only (cross-validation split)¶

import torch
from radharmony.dataset import RANZCRClipDataset

ds = RANZCRClipDataset(
    base_image_dir="~/.cache/kagglehub/competitions/ranzcr-clip-catheter-line-classification/",
    output_cls=True,
    include_test_split=False,
    dtype=torch.float32,
)
train_ds, val_ds = ds.get_datasets(n_splits=10)

Official train / test split¶

from radharmony.dataset import RANZCRClipDataset

ds = RANZCRClipDataset(
    base_image_dir="~/.cache/kagglehub/competitions/ranzcr-clip-catheter-line-classification/",
    output_cls=True,
    include_test_split=True,
)
train_ds, test_ds = ds.get_datasets_predefined()
# test_ds rows have NaN labels — use for inference/submission only

With polyline masks¶

from radharmony.dataset import RANZCRClipDataset

ds = RANZCRClipDataset(
    base_image_dir="~/.cache/kagglehub/competitions/ranzcr-clip-catheter-line-classification/",
    output_cls=True,
    output_mask=True,
    mask_output_dir="/tmp/ranzcr_masks/",   # rasterized PNGs written here
    mask_line_thickness=15,
)
train_ds, val_ds = ds.get_datasets(n_splits=10)
# sample["mask"] → torch.Tensor, shape (1, 224, 224)

Note

Polyline annotations only cover a subset of training images. Images without annotations will have NaN for mask_path and no "mask" key in their data dict.

Harmonizer¶

from radharmony.harmonizer import RANZCRClipHarmonizer

h = RANZCRClipHarmonizer(
    csv_path="~/.cache/kagglehub/competitions/ranzcr-clip-catheter-line-classification/train.csv",
    base_image_dir="~/.cache/kagglehub/competitions/ranzcr-clip-catheter-line-classification/",
    include_test_split=True,
)
df = h.harmonize()
print(df.columns.tolist())
print(df["split"].value_counts())
df.to_csv("ranzcr_clip_harmonized.csv", index=False)

Load from saved harmonized CSV¶

import pandas as pd
from radharmony.dataset import RANZCRClipDataset

ds = RANZCRClipDataset(
    base_image_dir="~/.cache/kagglehub/competitions/ranzcr-clip-catheter-line-classification/",
    harmonized_df=pd.read_csv("ranzcr_clip_harmonized.csv"),
    output_cls=True,
)

Outputs¶

Flag	Key	Shape	Notes
`output_cls=True`	`"cls"`	`(11,)`	Binary multi-label targets
`output_mask=True`	`"mask"`	`(1, 224, 224)`	Rasterized polyline mask; only for annotated images

Harmonizer notes¶

Raw CSV column names use "CVC - Abnormal" style; these are renamed to clean snake_case (cvc_abnormal) during harmonization
image_path is train/<StudyInstanceUID>.jpg or test/<StudyInstanceUID>.jpg, relative to base_image_dir
Test rows are discovered by walking <base_image_dir>/test/ for .jpg files; all label columns are NaN
A split column ("train" / "test") is always present in the harmonized DataFrame
Polyline masks are aggregated per study and JSON-encoded in mask_path; mask_output_dir triggers rasterization to PNG before MONAI loads them