jherng/rsna-2023-abdominal-trauma-detection

Dataset Card for RSNA 2023 Abdominal Trauma Detection (Preprocessed)

Dataset Summary

This dataset is the preprocessed version of the dataset from RSNA 2023 Abdominal Trauma Detection Kaggle Competition.

It is tailored for segmentation and classification tasks. It contains 3 different configs as described below:

• classification:
  • 4711 instances where each instance includes a CT scan in NIfTI format, target labels, and its relevant metadata.
• segmentation:
  • 206 instances where each instance includes a CT scan in NIfTI format, a segmentation mask in NIfTI format, and its relevant metadata.
• classification-with-mask:
  • 206 instances where each instance includes a CT scan in NIfTI format, a segmentation mask in NIfTI format, target labels, and its relevant metadata.

All CT scans and segmentation masks had already been resampled with voxel spacing (2.0, 2.0, 3.0) and thus its reduced file size.

Usage

from datasets import load_dataset

# Classification dataset
rsna_cls_ds = load_dataset("jherng/rsna-2023-abdominal-trauma-detection", streaming=True) # "classification" is the default configuration
rsna_cls_ds = load_dataset("jherng/rsna-2023-abdominal-trauma-detection", "classification", streaming=True) # download dataset on-demand and in-memory (no caching)
rsna_cls_ds = load_dataset("jherng/rsna-2023-abdominal-trauma-detection", "classification", streaming=False) # download dataset and cache locally (~90.09 GiB)
rsna_cls_ds = load_dataset("jherng/rsna-2023-abdominal-trauma-detection", "classification", streaming=True, test_size=0.05, random_state=42) # specify split size for train-test split

# Classification dataset with segmentation masks
rsna_clsmask_ds = load_dataset("jherng/rsna-2023-abdominal-trauma-detection", "classification-with-mask", streaming=True) # download dataset on-demand and in-memory (no caching)
rsna_clsmask_ds = load_dataset("jherng/rsna-2023-abdominal-trauma-detection", "classification-with-mask", streaming=False) # download dataset and cache locally (~3.91 GiB)
rsna_clsmask_ds = load_dataset("jherng/rsna-2023-abdominal-trauma-detection", "classification-with-mask", streaming=False, test_size=0.05, random_state=42) # specify split size for train-test split

# Segmentation dataset
rsna_seg_ds = load_dataset("jherng/rsna-2023-abdominal-trauma-detection", "segmentation", streaming=True) # download dataset on-demand and in-memory (no caching)
rsna_seg_ds = load_dataset("jherng/rsna-2023-abdominal-trauma-detection", "segmentation", streaming=False) # download dataset and cache locally (~3.91 GiB)
rsna_seg_ds = load_dataset("jherng/rsna-2023-abdominal-trauma-detection", "segmentation", streaming=True, test_size=0.1, random_state=42) # specify split size for train-test split

# Get the dataset splits
train_rsna_cls_ds = rsna_cls_ds["train"]; test_rsna_cls_ds = rsna_cls_ds["test"]
train_rsna_clsmask_ds = rsna_clsmask_ds["train"]; test_rsna_clsmask_ds = rsna_clsmask_ds["test"]
train_rsna_seg_ds = rsna_seg_ds["train"]; test_rsna_seg_ds = rsna_seg_ds["test"]

# Tip: Download speed up with multiprocessing
rsna_cls_ds = load_dataset("jherng/rsna-2023-abdominal-trauma-detection", streaming=False, num_proc=8) # num_proc: num of cpu core used for loading the dataset

Dataset Structure

Data Instances

Configuration 1: classification

• Size of downloaded dataset files: 90.09 GiB

An example of an instance in the 'classification' configuration looks as follows:

{
  "img_path": "https://huggingface.co/datasets/jherng/rsna-2023-abdominal-trauma-detection/resolve/main/train_images/25899/21872.nii.gz",
  "bowel": 0,
  "extravasation": 0,
  "kidney": 0,
  "liver": 0,
  "spleen": 0,
  "any_injury": false,
  "metadata": {
    "series_id": 21872,
    "patient_id": 25899,
    "incomplete_organ": false,
    "aortic_hu": 113.0,
    "pixel_representation": 0,
    "bits_allocated": 16,
    "bits_stored": 12
  }
}

Configuration 2: segmentation

• Size of downloaded dataset files: 3.91 GiB

An example of an instance in the 'segmentation' configuration looks as follows:

{
  "img_path": "https://huggingface.co/datasets/jherng/rsna-2023-abdominal-trauma-detection/resolve/main/train_images/4791/4622.nii.gz",
  "seg_path": "https://huggingface.co/datasets/jherng/rsna-2023-abdominal-trauma-detection/resolve/main/segmentations/4622.nii.gz",
  "metadata": {
    "series_id": 4622,
    "patient_id": 4791,
    "incomplete_organ": false,
    "aortic_hu": 223.0,
    "pixel_representation": 1,
    "bits_allocated": 16,
    "bits_stored": 16
  }
}

Configuration 3: classification-with-mask

• Size of downloaded dataset files: 3.91 GiB

An example of an instance in the 'classification-with-mask' configuration looks as follows:

{
  "img_path": "https://huggingface.co/datasets/jherng/rsna-2023-abdominal-trauma-detection/resolve/main/train_images/4791/4622.nii.gz",
  "seg_path": "https://huggingface.co/datasets/jherng/rsna-2023-abdominal-trauma-detection/resolve/main/segmentations/4622.nii.gz",
  "bowel": 0,
  "extravasation": 0,
  "kidney": 0,
  "liver": 1,
  "spleen": 1,
  "any_injury": true,
  "metadata": {
    "series_id": 4622,
    "patient_id": 4791,
    "incomplete_organ": false,
    "aortic_hu": 223.0,
    "pixel_representation": 1,
    "bits_allocated": 16,
    "bits_stored": 16
  }
}

Data Fields

The data fields for all configurations are as follows:

• img_path: a string feature representing the path to the CT scan in NIfTI format.
• seg_path: a string feature representing the path to the segmentation mask in NIfTI format (only for 'segmentation' and 'classification-with-mask' configurations).
• bowel, extravasation, kidney, liver, spleen: Class label features indicating the condition of respective organs.
• any_injury: a bool feature indicating the presence of any injury.
• metadata: a dictionary feature containing metadata information with the following fields:
  • series_id: an int32 feature.
  • patient_id: an int32 feature.
  • incomplete_organ: a bool feature.
  • aortic_hu: a float32 feature.
  • pixel_representation: an int32 feature.
  • bits_allocated: an int32 feature.
  • bits_stored: an int32 feature.

Data Splits

Default split:

• 0.9:0.1 with random_state = 42

Configuration Name	Train (n_samples)	Test (n_samples)
classification	4239	472
segmentation	185	21
classification-with-mask	185	21

Modify the split proportion:

rsna_ds = load_dataset("jherng/rsna-2023-abdominal-trauma-detection", "classification", test_size=0.05, random_state=42)

Additional Information

Citation Information

• Preprocessed dataset:

@InProceedings{huggingface:dataset,
title = {RSNA 2023 Abdominal Trauma Detection Dataset (Preprocessed)},
author={Hong Jia Herng},
year={2023}
}

• Original dataset:

@misc{rsna-2023-abdominal-trauma-detection,
    author = {Errol Colak, Hui-Ming Lin, Robyn Ball, Melissa Davis, Adam Flanders, Sabeena Jalal, Kirti Magudia, Brett Marinelli, Savvas Nicolaou, Luciano Prevedello, Jeff Rudie, George Shih, Maryam Vazirabad, John Mongan},
    title = {RSNA 2023 Abdominal Trauma Detection},
    publisher = {Kaggle},
    year = {2023},
    url = {https://kaggle.com/competitions/rsna-2023-abdominal-trauma-detection}
}