RSNA 2023 Abdominal Trauma¶

Modality: CT | Format: DICOM | Dim: 3D | Labels: 14 injury labels

Overview¶

RSNA 2023 Abdominal Trauma Detection is a Kaggle challenge dataset of 4,711 abdominal CT series from 3,147 patients (mean ~1.5 series per patient). Each series is a DICOM directory loaded as a 3-D volume. Fourteen binary labels cover bowel and extravasation injury (binary) plus kidney, liver, and spleen severity (3-class one-hot per organ) and an overall any_injury flag.

Patient-level labels are replicated across all series belonging to the same patient, matching the challenge framing ("predict patient-level injury from any of the patient's series").

Download¶

Available at Kaggle — RSNA 2023 Abdominal Trauma Detection.

Expected layout¶

rsna-2023-abdominal-trauma-detection/
  train_2024.csv
  train_series_meta.csv
  train_images/
    <patient_id>/
      <series_id>/
        <instance_number>.dcm
        ...

Label columns¶

Column	Description
`any_injury`	Any abdominal injury present
`bowel_healthy`	Bowel healthy
`bowel_injury`	Bowel injury
`extravasation_healthy`	No active extravasation
`extravasation_injury`	Active extravasation
`kidney_healthy`	Kidney healthy
`kidney_high`	Kidney high-grade injury
`kidney_low`	Kidney low-grade injury
`liver_healthy`	Liver healthy
`liver_high`	Liver high-grade injury
`liver_low`	Liver low-grade injury
`spleen_healthy`	Spleen healthy
`spleen_high`	Spleen high-grade injury
`spleen_low`	Spleen low-grade injury

The cls tensor has 14 values in alphabetical order.

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 `<patient_id>/<series_id>/` DICOM dirs (e.g. `~/datasets/external/rsna-2023-abdominal-trauma-detection/train_images/`)
`csv_path`	`str`	No	auto	`train_2024.csv`; auto-discovered in parent of `base_image_dir`
`series_meta_csv_path`	`str`	No	auto	`train_series_meta.csv`; auto-discovered
`hu_window`	`tuple` or `None`	No	`(-150, 250)`	HU clipping window (soft-tissue / contrast abdomen)

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`	Not supported in this release; silently ignored
`output_report`	`bool`	No	`False`	Not supported; silently ignored
`output_bbox`	`bool`	No	`False`	Not supported in this release; silently ignored
`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¶

from radharmony.dataset import RSNAAbdominalTrauma2023Dataset

ds = RSNAAbdominalTrauma2023Dataset(
    base_image_dir="/data/rsna-2023-abdominal-trauma-detection/train_images/",
    csv_path="/data/rsna-2023-abdominal-trauma-detection/train_2024.csv",
    series_meta_csv_path="/data/rsna-2023-abdominal-trauma-detection/train_series_meta.csv",
    output_cls=True,
    hu_window=(-150, 250),
)
train_ds, val_ds = ds.get_datasets(n_splits=5)

There is no separate test-split dataset class — the competition test images have no public labels. To load test volumes for inference, point base_image_dir at test_images/ and omit csv_path.

Harmonizer¶

from radharmony.harmonizer import RSNAAbdominalTrauma2023Harmonizer

h = RSNAAbdominalTrauma2023Harmonizer(
    csv_path="/data/rsna-2023-abdominal-trauma-detection/train_2024.csv",
    series_meta_csv_path="/data/rsna-2023-abdominal-trauma-detection/train_series_meta.csv",
    base_image_dir="/data/rsna-2023-abdominal-trauma-detection/train_images/",
)
df = h.harmonize()
print(df.columns.tolist())
df.to_csv("rsna_abdominal_trauma_harmonized.csv", index=False)

Load from saved harmonized CSV¶

import pandas as pd
from radharmony.dataset import RSNAAbdominalTrauma2023Dataset

ds = RSNAAbdominalTrauma2023Dataset(
    base_image_dir="/data/rsna-2023-abdominal-trauma-detection/train_images/",
    harmonized_df=pd.read_csv("rsna_abdominal_trauma_harmonized.csv"),
    output_cls=True,
)

Harmonizer notes¶

One row per series; patient-level labels from train_2024.csv are joined onto each series by patient_id
Series-level metadata aortic_hu (contrast-phase indicator) and incomplete_organ are carried through as extra columns in the harmonized DataFrame
image_path is <patient_id>/<series_id> (relative DICOM series directory); MONAI's ITKReader assembles per-slice .dcm files into a 3-D volume
HU window (-150, 250) covers soft-tissue / contrast-enhanced abdomen
bowel and extravasation are binary (healthy vs. injury); kidney, liver, spleen use 3-class one-hot columns
Organ segmentation masks and per-slice Active_Extravasation point annotations exist in the release but are not wired into this harmonizer

Outputs¶

Flag	Key	Shape	Notes
`output_cls=True`	`"cls"`	`(14,)`	Binary injury labels

Example paths¶

Role	Path
`base_image_dir` (train)	`/path/to/rsna-2023-abdominal-trauma-detection/train_images/`
`base_image_dir` (test)	`/path/to/rsna-2023-abdominal-trauma-detection/test_images/`
`csv_path` (`train_2024.csv`)	`/path/to/rsna-2023-abdominal-trauma-detection/train_2024.csv`
`series_meta_csv_path` (`train_series_meta.csv`)	`/path/to/rsna-2023-abdominal-trauma-detection/train_series_meta.csv`
test series meta (`test_series_meta.csv`)	`/path/to/rsna-2023-abdominal-trauma-detection/test_series_meta.csv`

Only one dataset class: RSNAAbdominalTrauma2023Dataset (no separate Train/Test class).