Overview

Schema Version

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

Name

Assessing the Performance of Models from the 2022 RSNA Cervical Spine Fracture Detection Competition at a Level I Trauma Center

Link

https://doi.org/10.1148/ryai.230550

Indexing

Keywords: Cervical spine fracture, CT, Emergency department, Machine learning, Convolutional Neural Network, Supervised learning, Genetic algorithms, Technology assessment, Head/Neck

Content: CT, ER, MK, NR

RadLex: RID10321, RID28769, RID28768, RID5362, RID25725

Author(s)

Zixuan Hu

Markand Patel

Robyn L. Ball

Hui Ming Lin

Luciano M. Prevedello

Mitra Naseri

Shobhit Mathur

Robert Moreland

Jefferson Wilson

Christopher Witiw

Kristen W. Yeom

Qishen Ha

Darragh Hanley

Selim Seferbekov

Hao Chen

Philipp Singer

Christof Henkel

Pascal Pfeiffer

Ian Pan

Harshit Sheoran

Wuqi Li

Adam E. Flanders

Felipe C. Kitamura

Tyler Richards

Jason Talbott

Ervin Sejdić

Errol Colak

Organization(s)

University of Toronto

St Michael’s Hospital, Unity Health Toronto

Edward S. Rogers Department of Electrical and Computer Engineering, University of Toronto

Department of Medical Imaging, Faculty of Medicine, University of Toronto

Li Ka Shing Knowledge Institute, Unity Health Toronto

The Jackson Laboratory

Stanford University, School of Medicine

H2O.ai

University of Birmingham, School of Computer Science

DoubleYard, Edulab Group

Mapbox

NVIDIA

Brigham and Women’s Hospital, Harvard Medical School

Goldsmiths, University of London

The Ohio State University, Department of Radiology

Thomas Jefferson University, Division of Neuroradiology

Universidade Federal de São Paulo (Unifesp)

University of California San Francisco

University of Utah, Department of Radiology and Imaging Sciences

North York General Hospital

Version

1.0

License

URL: https://www.ncbi.nlm.nih.gov/pmc/articles/PMC11605142

Contact

Corresponding author: Errol Colak (email as printed: ot.htlaeHytinU@kaloC.lorrE)

Funding

E.C. supported by the Odette Professorship in Artificial Intelligence for Medical Imaging, St Michael’s Hospital, Unity Health Toronto.

Ethical review

Retrospective IRB approval at Unity Health Toronto with waiver of informed consent.

Date

Published: 2024-09-18

References

[1] Hu Z, Patel M, Ball RL, et al.. "Assessing the Performance of Models from the 2022 RSNA Cervical Spine Fracture Detection Competition at a Level I Trauma Center". Radiology: Artificial Intelligence. 2024;6(6):e230550.. 2024-09-18. doi:10.1148/ryai.230550. PMID: 39298563. PMCID: PMC11605142.

Model

Architecture

Two-stage approach: segmentation (typically 2D or 3D U-Net) to isolate cervical spine voxels, followed by classification using CNN feature extraction, feature aggregation, and logits prediction.

Availability

Public competition dataset and award-winning models available at https://www.kaggle.com/competitions/rsna-2022-cervical-spine-fracture-detection

Clinical benefit

Assists in detection of cervical spine fractures on CT; potential to rapidly flag studies and identify some fractures missed by radiologists.

Clinical workflow phase

Clinical decision support; potential triage/flagging of studies in the ED setting.

Decision threshold

Per-model optimal thresholds determined on the public test set using Youden J: Qishen 0.72; Darragh 0.54; Selim 0.57; Speedrun 0.81; Skecherz 0.49; QWER 0.54; Harshit 0.72.

Degree of automation

Automated fracture probability prediction; intended as decision support for radiologists.

Indications for use

Adult emergency department patients undergoing cervical spine CT for traumatic indication in a level I trauma center environment; models trained on noncontrast CT and evaluated additionally on contrast-enhanced CT.

Input

Axial bone-window cervical spine CT series (≤1 mm section thickness); noncontrast and contrast-enhanced scans for testing (training data were noncontrast).

Instructions

Models produce examination-level fracture probabilities binarized using thresholds derived from the public test set via Youden J; applied to CT scans filtered to axial bone window images with ≤1 mm section thickness.

Limitations

Performance decreased on real-world clinical data compared to competition data, especially in contrast-enhanced scans; false positives associated with vessels, degenerative changes, osteophytes, calcifications, vascular channels, and artifacts; false negatives associated with degenerative changes, osteopenia, and certain fracture locations (end plate edges, transverse and spinous processes); single-center clinical dataset.

Output

CDEs: RDE2154, RDE2117

Description: Examination-level probability/classification of cervical spine fracture; visualization via Grad-CAM heat maps for interpretability; models developed to detect and localize fractures.

Recommendation

Models demonstrated high performance and identified some fractures missed by radiologists; warrant further evaluation for use as clinical support tools.

Reproducibility

Software stack detailed: Python 3.10.13; torch 2.1.0; SimpleITK 2.3.1; nibabel 5.2.0; torchvision 0.16.1; NumPy 1.26.2; scikit-image 0.22.0; opencv-python 4.8.1; pandas 2.1.4; matplotlib 3.8.0; Grad-CAM 1.4.8. Competition dataset and award-winning models available on Kaggle; private and clinical test sets not publicly available.

Use

Intended: Detection and diagnosis

User

Intended: Radiologist, Subspecialist diagnostic radiologist