CoolFace
Datasetpublic

MedOtter/COVID-19-20

COVID-19-20 Lung CT Lesion Segmentation Challenge Non-contrast chest CT with radiologist-verified binary COVID-19 lesion masks, from the MICCAI 2020 COVID-19 Lung CT Lesion Segmentation Challenge (a.k.a. COVID-19-20). ⚠️ Scope of this upload (training split only) This repository contains the public training split: 199 CT volumes, each with a ground-truth lesion mask. It is a faithful subset of the full challenge: Split Cases Masks Included here Train… See the full description on the dataset page: https://huggingface.co/datasets/MedOtter/COVID-19-20.

sourceHugging Facecc-by-4.0updated 3mo agoView on Hugging Face
0likes277downloads
Dataset Card

COVID-19-20 Lung CT Lesion Segmentation Challenge

Non-contrast chest CT with radiologist-verified binary COVID-19 lesion masks, from the MICCAI 2020 COVID-19 Lung CT Lesion Segmentation Challenge (a.k.a. COVID-19-20).

⚠️ Scope of this upload (training split only)

This repository contains the public training split: 199 CT volumes, each with a ground-truth lesion mask. It is a faithful subset of the full challenge:

SplitCasesMasksIncluded here
Train199✅ publicYes
Validation50❌ images-only (masks never released)No
Test46❌ fully withheldNo

The 50 validation cases ship as images only (no public masks) and the 46 test cases (23 Source A + 23 Source B) were never released — challenge evaluation ran server-side on Grand Challenge. Only the 199 training cases are usable for supervised segmentation, so only they are mirrored here.

Files

Train/
  volume-covid19-A-XXXX_ct.nii.gz    # CT volume, int16 Hounsfield units, (512, 512, Z)
  volume-covid19-A-XXXX_seg.nii.gz   # binary mask: 0 = background, 1 = COVID-19 lesion
COVID-19-20_TrainValidation.xlsx     # official train/validation filename lists

199 image+mask pairs, perfectly aligned (identical shape/affine). All cases are "Source A" (NIH multinational consortium). Patient IDs are unique — no patient appears twice.

Ground truth

Single annotation tier. Lesion masks were initialized by an NVIDIA+NIH deep model and then manually corrected by board-certified radiologists in ITK-SNAP for 3D-consistent ground truth. There are no competing rater / auto-vs-expert tiers to choose between.

License & provenance

  • —CT images: CC BY 4.0. Annotations: CC0.
  • —Organizer release (Roth et al., NIH / NVIDIA / Children's National) via the Grand Challenge platform. Source images originate from the TCIA collection CT Images in COVID-19 (DOI 10.7937/TCIA.2020.GQRY-NC81).
  • —Re-hosting permitted: CC BY 4.0 (images) + CC0 (masks) allow redistribution with attribution.

Relationship to other COVID CT datasets

No patient or source-archive overlap with the separate covid19-ct-seg (Ma et al. 20-case benchmark, sourced from Coronacases.org + Radiopaedia). Despite the confusingly similar "COVID-19-20-CTSEG" label some lists use for the Ma set, the two datasets are distinct and share no cases. COVID-19-20 derives from the TCIA CT Images in COVID-19 collection.

Citation

Roth HR, Xu Z, Tor-Díez C, et al. "Rapid artificial intelligence solutions in a pandemic — The COVID-19-20 Lung CT Lesion Segmentation Challenge." Medical Image Analysis 82:102605, 2022. doi:10.1016/j.media.2022.102605