Overview

Schema Version

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

Name

Precise Image-level CT Localization of Intracranial Hemorrhage with Models Trained on Study-level Labels

Link

https://dx.doi.org/10.1148/ryai.230296

Indexing

Keywords: intracranial hemorrhage, head CT, weakly supervised learning, attention-based bidirectional LSTM, EfficientNet-B2, transfer learning, Grad-CAM, study-level labels, image-level localization, worklist prioritization

Content: CT, NR, ER

RadLex: RID28768, RID4710, RID10321

SNOMED: 1386000

Author(s)

Yunan Wu

Michael Iorga

Suvarna Badhe

James Zhang

Donald R. Cantrell

Elaine J. Tanhehco

Nicholas Szrama

Andrew M. Naidech

Michael Drakopoulos

Shamis T. Hasan

Kunal M. Patel

Tarek A. Hijaz

Eric J. Russell

Shamal Lalvani

Amit Adate

Todd B. Parrish

Aggelos K. Katsaggelos

Virginia B. Hill

Organization(s)

Northwestern University

Northwestern University Feinberg School of Medicine

Shirley Ryan AbilityLab

Indiana University Health

Medical College of Wisconsin

McMaster University

Mount Sinai Medical Center

Version

1.0

License

Contact

gro.mn@llih.ainigriv

Funding

Supported by the fiscal year 2020 Dixon Translational Research Grants Initiative and the Northwestern Memorial Foundation project “Using Artificial Intelligence to Detect and Triage Emergencies on Noncontrast Head CTs.”

Ethical review

Institutional Review Board approval obtained with waivers of HIPAA authorization and informed consent.

Date

Updated: 2024-08-16

Published: 2024-08-28

Created: 2023-07-27

References

[1] Wu Y, Iorga M, Badhe S, Zhang J, Cantrell DR, Tanhehco EJ, Szrama N, Naidech AM, Drakopoulos M, Hasan ST, Patel KM, Hijaz TA, Russell EJ, Lalvani S, Adate A, Parrish TB, Katsaggelos AK, Hill VB. "Precise Image-level Localization of Intracranial Hemorrhage on Head CT Scans with Deep Learning Models Trained on Study-level Labels". Radiology: Artificial Intelligence. 2024;6(6):e230296.. 2024. doi:10.1148/ryai.230296. PMID: 39194400. PMCID: PMC11605431.

Model

Architecture

End-to-end pipeline: EfficientNet-B2 CNN pretrained on RSNA images (stage I) + attention-based bidirectional LSTM trained at study level (stage II) + fine-tuning of the entire network on local data (stage III); Grad-CAM used for interpretability.

Availability

Source code and model weights available: https://github.com/YunanWu2168/Image-level-Localization-of-ICH-Head-CTs-with-DL-Models-Trained-on-Study-level-Labels

Clinical benefit

Automated detection and prioritization of intracranial hemorrhage on noncontrast head CT to expedite reading, improve triage, and reduce false-positive interruptions.

Clinical workflow phase

patients’ triage; workflow optimization; clinical decision support systems

Decision threshold

Degree of automation

Assists radiologists by providing study-level ICH probabilities, slice-level attention scores, and heatmaps for localization; not autonomous.

Indications for use

Detection and localization of intracranial hemorrhage on noncontrast head CT scans at the study and image levels in a radiology reading environment, including emergency settings for worklist prioritization.

Input

Noncontrast head CT series (DICOM), study-level labels derived via NLP for training; RSNA dataset for pretraining and CQ500 for external testing.

Instructions

Model provides study-level ICH probability, slice-level attention weights (0–1) for image-level localization, and Grad-CAM heatmaps; trained via three stages (RSNA pretraining, study-level LSTM with attention, and end-to-end fine-tuning). Ensemble of five folds can be used to improve robustness.

Limitations

Study-level labels derived from radiology reports using NLP may introduce labeling errors; Grad-CAM saliency maps may have limited robustness; some subtle or trace hemorrhages and artifact-affected cases were misclassified; prospective performance with radiologist assistance not yet assessed; regulatory clearance not obtained.

Output

CDEs: RDE1288, RDE1287, RDE2267

Description: Study-level probability of intracranial hemorrhage; image-level attention scores indicating likelihood of ICH on each image; Grad-CAM heatmaps localizing suspected ICH regions; optional binary masks after thresholding heatmaps for visualization.

Recommendation

Use as a triage/alert and prioritization tool to flag likely ICH studies and guide image-level review; radiologist oversight required.

Regulatory information

Comment: This is an academic research model reported in Radiology: Artificial Intelligence; no regulatory submission reported.

Authorization status: Not FDA cleared/authorized (research study)

Reproducibility

Fivefold cross-validation on local dataset with a fixed held-out test set; external validation on CQ500; code and pretrained weights publicly available for replication.

Sustainability

Use

Intended: detection, triage

User

Intended: Radiologist

Excluded: Layperson