Morris-is-taken/RSNA_23_256x256_ROI_PNG

RSNA 2023 Abdominal Trauma Detection

This dataset is a preprocessed version of the RSNA 2023 Abdominal Trauma Detection training set.

Each DICOM scan has been converted into 96 PNG slices of fixed size 256×256:

• Scans with fewer slices were upsampled.
• Scans with more slices were downsampled.

The dataset is organized into 5 folds for cross-validation.

Contents

• arrow_cache/: Arrow cache of training/validation set (fold 1).
• fold_*.tar: Preprocessed images grouped by patient/series/organ.
• RSNA_23_256x256_ROI_PNG.py: Dataset loading script for Hugging Face datasets.

Usage

from datasets import load_dataset

# Load Fold 1
ds = load_dataset(
    "Morris-is-taken/RSNA_23_256x256_ROI_PNG",
    name="fold1",
    trust_remote_code=True
)

print(ds["train"])
# Dataset({
#     features: ['patient_id', 'series_id', 'organs', 'image', 'img_idx'],
#     num_rows: 992468
# })
print(ds["validation"])
# Dataset({
#     features: ['patient_id', 'series_id', 'organs', 'image', 'img_idx'],
#     num_rows: 248117
# })

Dataset features

• patient_id: Unique identifier for each patient.
• series_id: Identifier for the scanning series. A patient may have one or more series.
• organs: Target region of interest
  • 0: Full image
  • 1: Kidney
  • 2: Liver
  • 3: Spleen
• image: A PIL Image object representing the slice.
• img_idx: Index of the slice within the series.

Requirement

pip install datasets==3.6.0