Overview

Schema Version

https://atlas.rsna.org/schemas/2025-11/dataset.json

Name

Atlantis-labeled head CT dataset (UCSF, 2010-2017)

Link

https://pmc.ncbi.nlm.nih.gov/articles/PMC11140498

Indexing

Keywords: intracranial hemorrhage, head CT, segmentation, semi-supervised learning, traumatic brain injury

Content: CT, NR

RadLex: RID45978, RID5143, RID10782, RID10321, RID4710

SNOMED: 1386000

Author(s)

Emily Lin

Esther L. Yuh

Organization(s)

University of California San Francisco, Department of Radiology & Biomedical Imaging

Contact

Corresponding author: Esther L. Yuh (email as printed in article: ude.fscu@huy.rehtse)

Funding

Supported by the California Institute to Advance Precision Medicine (CIAPM).

Ethical review

Retrospective study approved by the University of California San Francisco IRB; HIPAA compliant; waiver of consent per 45 CFR 46.116(d).

Comments

Pixel-labeled clinical head CT dataset used to train the teacher model for semi-supervised learning of intracranial hemorrhage detection/segmentation.

References

[1] Lin E, Yuh EL. "Semi-supervised Learning for Generalizable Intracranial Hemorrhage Detection and Segmentation". Radiology: Artificial Intelligence. 2024-03-06. doi:10.1148/ryai.230077. PMID: 38446043. PMCID: PMC11140498.

[2] Kuo W, Hӓne C, Mukherjee P, Malik J, Yuh EL. "Expert-level detection of acute intracranial hemorrhage on head computed tomography using deep learning". Proc Natl Acad Sci U S A. 2019. doi:10.1073/pnas.1908021116. PMID: 31636195. PMCID: PMC6842581.

Dataset

Motivation

Provide pixel-level labels for intracranial hemorrhage to train and evaluate deep learning models and to bootstrap semi-supervised learning for improved generalizability.

Sampling

Single institution, four 64–detector row CT scanners from a single vendor; scans from 2010–2017.

Partitioning scheme

Used as the labeled training set (teacher model) within the study; validation and test sets came from separate sources.

Missing information

Public access details, exact number of patients, image acquisition parameters, and file resolution not provided.

Noise

Includes typical clinical CT artifacts and low SNR seen in practice.

External data

Used in combination with a large unlabeled external corpus (RSNA/ASNR Kaggle Intracranial Hemorrhage Detection) for semi-supervised training.

Confidentiality

De-identified; skull, scalp, and face removed for anonymity.

Re-identification

Faces/skull/scalp removed to reduce re-identification risk.

Sensitive data

Clinical imaging data; de-identified with anatomic facial structures removed.