RICORD — 2D Slice-Based Chest CT Dataset with COVID-19 Labels

Axial chest CT slices from the RSNA International COVID-19 Open Radiology Database (RICORD), releases 1a (COVID-19 positive) and 1b (COVID-19 negative), converted to a 2D, slice-level dataset by Chehab Lab. Each row is one slice with the patient's COVID-19 label and the available acquisition and patient metadata.

📦 Dataset Structure

• 227 patients, 227 CT series, 40,104 slices. One series per patient.
• Label: covid, with 117 negative (RICORD 1b) and 110 positive (RICORD 1a) patients.

column	type	description
volume_id	string	RICORD patient ID
series_uid	string	DICOM SeriesInstanceUID of the selected series
slice_id	int32	axial slice index in the volume (inferior → superior)
image	image	CT slice, 8-bit grayscale PNG
covid	int32	1 = COVID-19 positive (1a), 0 = negative (1b)
collection	string	MIDRC-RICORD-1A or MIDRC-RICORD-1B
slice_thickness	float32	slice spacing (mm)
pixel_spacing	float32	in-plane pixel spacing (mm)
manufacturer	string	scanner manufacturer
kernel	string	reconstruction kernel
patient_age	float32	age in years (from the DICOM header)
patient_sex	string	sex (from the DICOM header)

⚙️ Preprocessing

• Data downloaded from The Cancer Imaging Archive (TCIA) with the NBIA API.
• One CT series per patient: the series with the most images (RICORD 1a often has two reconstructions of the same exam).
• DICOM slices are sorted by their physical position. Slices whose size differs from the rest of the series are dropped.
• Pixel values are converted to Hounsfield units and windowed to the lung window [−1000, 400] HU, then scaled to 0–255 and saved as 8-bit PNG.
• The original in-plane resolution is kept with no resizing. Most volumes are 384×384 (751), 640×640 (403), 1024×1024 (216) or 320×320 (83); a few have other sizes. Resize before training if needed.
• Constant (empty) slices are removed. slice_id keeps the index of the slice in the volume.

🚀 Usage

from datasets import load_dataset
import matplotlib.pyplot as plt

ds = load_dataset("chehablab/RICORD", split="train")
sample = ds[100]

plt.imshow(sample["image"], cmap="gray")
plt.title(f"{sample['volume_id']} | slice {sample['slice_id']} | covid={sample['covid']}")
plt.axis("off")
plt.show()

🏷️ Label Descriptions

COVID-19 (covid)

• 1 (RICORD 1a): chest CT of adult patients (≥ 18 years) with a positive COVID-19 diagnosis.
• 0 (RICORD 1b): chest CT of adult patients (≥ 18 years) with a negative COVID-19 diagnosis.

🧪 Use Cases

• Patient-level COVID-19 classification from 3D chest CT (multiple instance learning, slice aggregation)
• Slice-level representation learning and benchmarking of pretrained encoders on a diffuse lung finding
• Studying the effect of acquisition parameters (slice thickness, kernel, manufacturer)

⚠️ Important Notes

• Split by volume_id, not by row, so slices of the same patient are never in both train and test.
• covid is a patient-level label: every slice of a patient has the same label, even slices without visible disease.
• The positive and negative classes come from two different releases (1a and 1b). Acquisition differences between them (for example slice thickness, kernel, site) can act as a shortcut, so check models for this confound.
• patient_age is missing for about 21% of patients because it was removed during de-identification.
• The expert annotations released with RICORD 1a (segmentations and exam-level labels) are not included.

📚 Citation

If you use this dataset, please cite Chehab Lab and the original RICORD work:

@article{tsai2021ricord,
  title   = {The RSNA International COVID-19 Open Radiology Database (RICORD)},
  author  = {Tsai, Emily B. and Simpson, Scott and Lungren, Matthew P. and others},
  journal = {Radiology},
  volume  = {299},
  number  = {1},
  pages   = {E204--E213},
  year    = {2021},
  doi     = {10.1148/radiol.2021203957}
}

Please also cite The Cancer Imaging Archive: Clark K. et al., "The Cancer Imaging Archive (TCIA): Maintaining and Operating a Public Information Repository", Journal of Digital Imaging, 26(6), 1045–1057, 2013.

Original sources: MIDRC-RICORD-1a · MIDRC-RICORD-1b

📜 License

CC BY-NC 4.0, the same as the original RICORD collections: attribution is required and commercial use is not permitted.

Chehab Lab @ 2026