RSNA 2022 Cervical Spine

Modality: CT | Format: DICOM | Dim: 3D | Labels: 8 fracture labels

Overview

RSNA 2022 Cervical Spine Fracture Detection is a Kaggle challenge dataset containing 2,019 cervical-spine CT studies (one per patient). Each study is a directory of per-slice DICOM files loaded as a 3-D volume by MONAI's ITKReader. Eight study-level binary labels encode presence/absence of fracture at each vertebra (C1–C7) and an overall flag (patient_overall).

A subset of 87 studies ships with whole-volume NIfTI segmentation masks. A separate bbox variant (RSNA2022CervicalSpineBboxDataset) provides per-volume 3-D fracture bounding boxes for the 235 studies that have at least one fracture-localizing annotation.

Download

Available at Kaggle — RSNA 2022 Cervical Spine Fracture Detection.

Expected layout

rsna-2022-cervical-spine-fracture-detection/
  train.csv
  train_bounding_boxes.csv
  segmentations/
    <StudyInstanceUID>.nii
    ...
  train_images/
    <StudyInstanceUID>/
      100.dcm
      101.dcm
      ...

Label columns

Column	Description
c1	C1 vertebra fracture
c2	C2 vertebra fracture
c3	C3 vertebra fracture
c4	C4 vertebra fracture
c5	C5 vertebra fracture
c6	C6 vertebra fracture
c7	C7 vertebra fracture
patient_overall	Any cervical fracture present

The cls tensor has values in sorted (alphabetical) order: c1, c2, c3, c4, c5, c6, c7, patient_overall.

---

3D Volume variant — RSNA2022CervicalSpineDataset

Constructor arguments

Argument	Type	Required	Default	Description
base_image_dir	str	Yes*	None	train_images/ (or test_images/ for the test split) — direct parent of <StudyInstanceUID>/ DICOM dirs (e.g. ~/datasets/external/rsna-2022-cervical-spine-fracture-detection/train_images/)
csv_path	str	No	auto	train.csv; auto-discovered in parent of base_image_dir
segmentation_dir	str	No	auto	Directory of <StudyUID>.nii masks; auto-discovered as ../segmentations/
hu_window	tuple or None	No	(-200, 1800)	HU clipping window (bone-centric default)

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"; restricts dataset to 87 studies with NIfTI masks
output_report	bool	No	False	Not supported; silently ignored
output_bbox	bool	No	False	Not supported; use RSNA2022CervicalSpineBboxDataset for bboxes
transform	MONAI Compose	No	3D pipeline	Custom MONAI transform
cache_dir	str	No	"./cache"	MONAI PersistentDataset cache; None disables caching
dtype	torch.dtype	No	torch.bfloat16	Output tensor dtype; use torch.float32 on CPU
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

*base_image_dir can be omitted when harmonizer_path, harmonizer, or harmonized_df is provided.

Dataset constructor

The constructor arguments above apply to both RSNA2022CervicalSpineTrainDataset and RSNA2022CervicalSpineTestDataset. The test split has no labels or segmentations — output_cls, output_mask, and output_bbox are silently ignored.

Train split

import torch
from radharmony.dataset import RSNA2022CervicalSpineTrainDataset

ds = RSNA2022CervicalSpineTrainDataset(
    base_image_dir="/data/rsna-2022-cervical-spine-fracture-detection/train_images/",
    csv_path="/data/rsna-2022-cervical-spine-fracture-detection/train.csv",
    output_cls=True,
    hu_window=(-200, 1800),
)
train_ds, val_ds = ds.get_datasets(n_splits=5)

With segmentation masks (restricts to 87 studies):

ds = RSNA2022CervicalSpineTrainDataset(
    base_image_dir="/data/rsna-2022-cervical-spine-fracture-detection/train_images/",
    segmentation_dir="/data/rsna-2022-cervical-spine-fracture-detection/segmentations/",
    output_cls=True,
    output_mask=True,
)

Test split

from radharmony.dataset import RSNA2022CervicalSpineTestDataset

ds = RSNA2022CervicalSpineTestDataset(
    base_image_dir="/data/rsna-2022-cervical-spine-fracture-detection/test_images/",
    hu_window=(-200, 1800),
)

Harmonizer: instantiate and inspect

from radharmony.harmonizer import RSNA2022CervicalSpineTrainHarmonizer

h = RSNA2022CervicalSpineTrainHarmonizer(
    csv_path="/data/rsna-2022-cervical-spine-fracture-detection/train.csv",
    base_image_dir="/data/rsna-2022-cervical-spine-fracture-detection/train_images/",
    segmentation_dir="/data/rsna-2022-cervical-spine-fracture-detection/segmentations/",
)
df = h.harmonize()
print(df.columns.tolist())
df.to_csv("rsna_cervical_spine_harmonized.csv", index=False)

Load from saved harmonized CSV

import pandas as pd
from radharmony.dataset import RSNA2022CervicalSpineTrainDataset

ds = RSNA2022CervicalSpineTrainDataset(
    base_image_dir="/data/rsna-2022-cervical-spine-fracture-detection/train_images/",
    harmonized_df=pd.read_csv("rsna_cervical_spine_harmonized.csv"),
    output_cls=True,
)

Harmonizer notes

• One row per StudyInstanceUID (= one CT volume)
• image_path is a DICOM study directory (single level of nesting under train_images/)
• MONAI's ITKReader assembles the per-slice .dcm files into a 3-D volume
• mask_path is populated for the 87 studies that have a matching .nii file; NaN for the rest
• output_mask=True naturally restricts the dataset to the 87 mask-having studies via the framework's dropna on mask_path
• HU window (-200, 1800) is bone-centric; use (-1000, 1000) for general soft-tissue
• CSV columns C1..C7 (uppercase) are renamed to lowercase c1..c7 by the harmonizer

Outputs

Flag	Key	Shape	Notes
output_cls=True	"cls"	(8,)	Binary fracture labels
output_mask=True	"mask"	(1, D, H, W)	NIfTI segmentation (87 studies only)

---

3D Bbox variant — RSNA2022CervicalSpineBboxDataset

Constructor arguments

Argument	Type	Required	Default	Description
base_image_dir	str	Yes*	None	train_images/ directory — direct parent of <StudyInstanceUID>/ DICOM dirs (e.g. ~/datasets/external/rsna-2022-cervical-spine-fracture-detection/train_images/)
csv_path	str	No	auto	train.csv; auto-discovered
bbox_csv_path	str	No	auto	train_bounding_boxes.csv; auto-discovered

Shared arguments (inherited from BaseRadiologicalDataset)

Argument	Type	Required	Default	Description
output_cls	bool	No	False	Include study-level "cls" (8-D) in data dict
output_mask	bool	No	False	Not supported; silently ignored
output_report	bool	No	False	Not supported; silently ignored
output_bbox	bool	No	False	Include per-volume "bbox" and "bbox_labels"
transform	MONAI Compose	No	3D pipeline	Custom MONAI transform
cache_dir	str	No	"./cache"	MONAI PersistentDataset cache; None disables
dtype	torch.dtype	No	torch.bfloat16	Output tensor dtype; use torch.float32 on CPU
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

*base_image_dir can be omitted when harmonizer_path, harmonizer, or harmonized_df is provided.

Dataset constructor

from radharmony.dataset import RSNA2022CervicalSpineBboxDataset

ds = RSNA2022CervicalSpineBboxDataset(
    base_image_dir="/data/rsna-2022-cervical-spine-fracture-detection/train_images/",
    csv_path="/data/rsna-2022-cervical-spine-fracture-detection/train.csv",
    bbox_csv_path="/data/rsna-2022-cervical-spine-fracture-detection/train_bounding_boxes.csv",
    output_cls=True,
    output_bbox=True,
)
train_ds, val_ds = ds.get_datasets(n_splits=5)

Harmonizer: instantiate and inspect

from radharmony.harmonizer import RSNA2022CervicalSpineBboxHarmonizer

h = RSNA2022CervicalSpineBboxHarmonizer(
    csv_path="/data/rsna-2022-cervical-spine-fracture-detection/train.csv",
    bbox_csv_path="/data/rsna-2022-cervical-spine-fracture-detection/train_bounding_boxes.csv",
    base_image_dir="/data/rsna-2022-cervical-spine-fracture-detection/train_images/",
)
df = h.harmonize()
print(df.columns.tolist())
df.to_csv("rsna_cervical_spine_bbox_harmonized.csv", index=False)

Load from saved harmonized CSV

import pandas as pd
from radharmony.dataset import RSNA2022CervicalSpineBboxDataset

ds = RSNA2022CervicalSpineBboxDataset(
    base_image_dir="/data/rsna-2022-cervical-spine-fracture-detection/train_images/",
    harmonized_df=pd.read_csv("rsna_cervical_spine_bbox_harmonized.csv"),
    output_cls=True,
    output_bbox=True,
)

Harmonizer notes

• One row per StudyInstanceUID (= one CT volume) with at least one fracture bbox
• image_path is the <StudyUID>/ DICOM directory; ITKReader assembles it into a 3-D volume
• Multiple fracture annotations are aggregated into a list of [d_min, d_max, y_min, y_max, x_min, x_max] 6-tuples (post-transpose (D, H, W) axis order)
• Coordinates are normalised to [0, 1] against the loaded volume dimensions
• Study-level cls labels are joined from train.csv by StudyInstanceUID

Outputs

Flag	Key	Shape	Notes
output_cls=True	"cls"	(8,)	Study-level binary fracture labels
output_bbox=True	"bbox"	list of [d_min, d_max, y_min, y_max, x_min, x_max]	Fractional coords (post-transpose (D, H, W) order); one entry per fracture annotation in the study
output_bbox=True	"bbox_labels"	list of str	Label name per bounding box

Example paths

Role	Path
base_image_dir (train)	/path/to/rsna-2022-cervical-spine-fracture-detection/train_images/
base_image_dir (test)	/path/to/rsna-2022-cervical-spine-fracture-detection/test_images/
csv_path (train.csv)	/path/to/rsna-2022-cervical-spine-fracture-detection/train.csv
bbox_csv_path (train_bounding_boxes.csv)	/path/to/rsna-2022-cervical-spine-fracture-detection/train_bounding_boxes.csv