MIDRC-RICORD-1a — COVID-19 Chest CT Lesion Segmentation

Non-contrast chest CT of COVID-19-positive patients with multi-class airspace-disease segmentation masks, from the MIDRC RSNA International COVID-19 Open Radiology Database (RICORD), Release 1a.

⚠️ This is a DERIVED / reformatted dataset

The official TCIA MIDRC-RICORD-1a archive is imaging-only DICOM with no masks. Segmentation is distributed separately as an MD.ai JSON of per-slice freeform polygons (keyed by SOPInstanceUID). This repository is built by:

1. Converting each annotated DICOM series to a NIfTI volume (int16 Hounsfield units, RescaleSlope/Intercept applied).
2. Rasterizing the MD.ai polygons (cv2.fillPoly) onto each slice to produce a volumetric, multi-class mask co-registered to the image (identical shape/affine).

It is not a byte-for-byte mirror of the TCIA DICOM. For the raw DICOM use TCIA directly (DOI 10.7937/VTW4-X588).

Multi-rater ground truth (UNION policy)

RICORD-1a was annotated by three teams ("Purple Group 1", "Team 1", "Team 2") that mostly partitioned the cohort — each segmented a different set of studies, with only 12 series double-read for inter-rater overlap. No single team covers the whole cohort.

This release uses the UNION of all three teams as the ground truth: one mask per series = the logical OR of every team's polygons. For the 12 double-read series the polygons are merged. Where polygons of different classes overlap on a slice, a fixed paint order lets the more specific/rarer finding win (opacity < other < atelectasis < TIB < nodule < cavity).

metadata.csv records, per series: annotator_groups, n_annotators, is_double_read — so the 12 double-read cases and each team's contribution remain identifiable.

Coverage

Patients (cohort)	110 (all COVID-19 positive); 108 have a segmented series
Studies / series segmented	118 (one diagnostic axial series per study; 2 of 120 studies — 2 patients — have no segmentation)
Total CT series on TCIA	229 (reformats / scouts / series without masks are not included)
Segmented slices	9,166 (only diseased slices carry polygons; the rest are background)
Double-read series	12 (flagged is_double_read=1; polygons unioned)

Label map (class_map.json)

id	class	note
0	background
1	Infectious opacity	dominant COVID finding (ground-glass / consolidation)
2	Infectious TIB/micronodules	sparse
3	Infectious cavity	sparse
4	Noninfectious nodule/mass	sparse
5	Atelectasis	sparse
6	Other noninfectious opacity	sparse

The mask is overwhelmingly class 1 (Infectious opacity); classes 2–6 are a sparse long tail. For a binary lesion task, treat any value > 0 as foreground.

Files

images/<patient_id>.nii.gz   # CT volume, int16 HU, [row, col, slice]
masks/<patient_id>.nii.gz    # uint8 multi-class mask {0..6}, co-registered
metadata.csv                 # per-series: UIDs, annotator groups, demographics, PCR, per-class voxels
class_map.json               # {class_id: name}

metadata.csv preserves cross-reference IDs (patient_id, study_uid, series_uid, site) plus joined clinical data from the official RICORD-1a clinical spreadsheet (age, sex, exam_description, pcr_result, specimen_source) and the exam-level COVID appearance category (covid_appearance: Typical / Indeterminate / Atypical / Negative for pneumonia).

License & provenance

• License: CC BY-NC 4.0 (non-commercial). Official source: TCIA / MIDRC (RSNA). Not a re-host.
• Data DOI: 10.7937/VTW4-X588.

Relationship to other datasets

• Sibling MIDRC collections RICORD-1b (COVID-negative CT) and RICORD-1c (COVID-positive chest radiographs) are disjoint cohorts, same provenance.
• No patient or source-archive overlap with the suite's other COVID-CT segmentation sets (covid19-ct-seg / Ma 20-case; COVID-19-20 / Roth) — those derive from independent (Coronacases/Radiopaedia and the TCIA "CT Images in COVID-19") sources. The site column (de-identified site token from the patient ID) can be used for any future de-duplication.

Citation

Tsai EB, Simpson S, Lungren MP, et al. "The RSNA International COVID-19 Open Radiology Database (RICORD)." Radiology 2021;299(1):E204–E213. doi:10.1148/radiol.2021203957

Data: Tsai E, Simpson S, Lungren MP, et al. (2020). MIDRC-RICORD-1a [Data set]. The Cancer Imaging Archive. https://doi.org/10.7937/VTW4-X588