Overview

Schema Version

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

Name

RadImageNet pretrained convolutional neural networks (CNNs)

Link

https://github.com/BMEII-AI/RadImageNet

Indexing

Keywords: RadImageNet, transfer learning, pretrained models, medical imaging, CT, MRI, ultrasound, dataset, Grad-CAM

Content: CT, MR, US, CH, NR, MK, GI, OI, IN

SNOMED: 237495005, 1386000, 840539006, 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)

BioMedical Engineering and Imaging Institute, Icahn School of Medicine at Mount Sinai

Department of Diagnostic, Interventional and Molecular Radiology, Icahn School of Medicine at Mount Sinai

Department of Mathematics, University of Oklahoma

Department of Radiology, Weill Cornell Medicine

Department of Radiology, East River Medical Imaging

Version

1.0

License

Text: CC BY 4.0

URL: https://creativecommons.org/licenses/by/4.0/

Contact

Yang Yang; email: ude.ainigriv@cc5yy

Funding

Authors declared no funding for this work.

Ethical review

Institutional review boards waived the requirement for written informed consent for this retrospective, HIPAA-compliant study using de-identified data.

Date

Updated: 2022-09-01

Published: 2022-07-27

Created: 2021-12-17

References

[1] Mei X, Liu Z, Robson PM, et al.. "RadImageNet: An Open Radiologic Deep Learning Research Dataset for Effective Transfer Learning". Radiology: Artificial Intelligence. 2022 Sep;4(5):e210315.. 2022-07-27. doi:10.1148/ryai.210315. PMID: 36204533. PMCID: PMC9530758.

Model

Architecture

Convolutional Neural Networks: Inception-ResNet-v2, ResNet50, DenseNet121, and InceptionV3 trained from scratch on RadImageNet with randomly initialized weights.

Availability

Pretrained models and code: https://github.com/BMEII-AI/RadImageNet; Data access requests: http://radimagenet.com

Clinical benefit

Provides improved starting weights for transfer learning in radiologic AI applications, especially for small datasets, yielding higher AUCs and more consistent, localized attention maps compared with ImageNet pretraining.

Degree of automation

Automates feature extraction and provides pretrained weights for downstream model development; not a standalone diagnostic device.

Indications for use

Research use for initializing deep learning models in medical imaging tasks (classification and segmentation) across CT, MRI, and ultrasound domains.

Input

Grayscale CT, MRI, and ultrasound key images resized to 224×224 pixels for pretraining (RadImageNet) and 256×256 pixels for downstream tasks.

Instructions

For fine-tuning, unfreezing all layers consistently achieved best performance; a smaller learning rate (e.g., 0.0001) is suggested when training all layers. Add global average pooling, dropout (0.5), and softmax output for classification; use patient-wise data splits.

Limitations

Single-image key findings may not reflect full clinical workflow; some images contain multiple findings but only one label was used; ROIs defined during clinical interpretation were not used in training; reduced-resolution images may obscure small findings; 165 categories grouped by ICD-10 and imaging characteristics are not diagnostic; no radiography images in the RadImageNet pretraining dataset; number of classes fewer than ImageNet.

Output

CDEs: RDE205, RDE746, RDE226

Description: During pretraining, models output probabilities across 165 pathologic labels (softmax). When transferred, outputs include task-specific classification scores and Grad-CAM localization maps; Dice used when segmentation ground truth available.

Recommendation

Use RadImageNet-pretrained weights as initialization for medical imaging tasks, particularly when target datasets are small or modality/anatomic region overlaps with CT, MRI, or ultrasound.

Regulatory information

Comment: Research-only pretrained models and dataset; no regulatory clearance claimed.

Reproducibility

Training-testing splits were patient-wise; multiple architectures and 24 fine-tuning scenarios reported; code and pretrained weights are publicly available.

Use

Intended: Other

Out-of-scope: Diagnosis

Excluded: Other

User

Intended: Researcher

Out-of-scope: Patient, Layperson

Excluded: Referring provider