Overview

Schema Version

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

Name

RadImageNet

Link

http://radimagenet.com

Indexing

Keywords: RadImageNet, transfer learning, pretrained models, CT, MRI, ultrasound, radiology dataset, 1.35 million images, 165 labels, key images

Content: CH, MK, NR, GI, OI, OT

RadLex: RID10326, RID10312, RID10321

SNOMED: 882784691000119000, 237495005, 1386000, 233604007

Author(s)

Xueyan Mei

Zelong Liu

Philip M. Robson

Brett Marinelli

Mingqian Huang

Amish Doshi

Adam Jacobi

Chendi Cao

Katherine E. Link

Thomas Yang

Ying Wang

Hayit Greenspan

Timothy Deyer

Zahi A. Fayad

Yang Yang

Organization(s)

Icahn School of Medicine at Mount Sinai

East River Medical Imaging

Weill Cornell Medicine (Department of Radiology)

University of Oklahoma (Department of Mathematics)

Funding

Authors declared no funding for this work.

Ethical review

Retrospective HIPAA-compliant study with IRB waiver of informed consent; de-identified data; third-party certification of de-identified data transfer; no link between patients, data provider, and data receiver was made available.

Comments

Large-scale medical imaging dataset constructed from clinical key images and labels retrospectively extracted from radiologist interpretations; includes normal studies; used to train and release pretrained CNN models for transfer learning.

Date

Published: 2022-07-27

References

[1] Mei X, Liu Z, Robson PM, Marinelli B, Huang M, Doshi A, Jacobi A, Cao C, Link KE, Yang T, Wang Y, Greenspan H, Deyer T, Fayad ZA, Yang Y.. "RadImageNet: An Open Radiologic Deep Learning Research Dataset for Effective Transfer Learning". Radiology: Artificial Intelligence. 2022-07-27. doi:10.1148/ryai.210315. PMID: 36204533. PMCID: PMC9530758.

Dataset

Motivation

Create a large-scale medical imaging-only dataset to enable effective transfer learning for diverse radiologic AI applications.

Sampling

Key images chosen by reading radiologists during routine clinical interpretation; 8,528 normal studies identified via report review included with all diagnostic sequences and images.

Partitioning scheme

Patient-level split: 75% training, 10% validation, 15% test within RadImageNet.

Missing information

File formats, demographic distributions, and per-partition counts are not specified in the article.

Relationships between instances

Multiple images per patient; radiologist-selected key images per study with single pathologic label and available region-of-interest; includes axial, sagittal, and coronal views and any sequence as long as pathology is clearly represented.

External data

Downstream evaluation used external public datasets for thyroid/breast ultrasound, knee MRI, chest radiography pneumonia, COVID-19 CT, SARS-CoV-2 CT, and intracranial hemorrhage CT (as cited in the article).

Confidentiality

De-identified data with third-party certification; no re-linkage between patients, data provider, and receiver.

Re-identification

Described as de-identified with no linkage; third-party certification of de-identified data transfer.

Sensitive data

Patient identifiers removed; HIPAA-compliant handling as described.