Database Credentialed Access

Fracture Detection on Cervical Spine CT

Sebnem Kuzulugil Muhammad Mamdani Hui-Ming Lin Priscila Crivellaro Aditya Bharatha Jefferson Wilson Monica Tafur Suradech Suthiphosuwan Oleksandra Samorodova Errol Colak

Published: Feb. 10, 2023. Version: 1.0.0 <View latest version>

When using this resource, please cite: (show more options)
Kuzulugil, S., Mamdani, M., Lin, H., Crivellaro, P., Bharatha, A., Wilson, J., Tafur, M., Suthiphosuwan, S., Samorodova, O., & Colak, E. (2023). Fracture Detection on Cervical Spine CT (version 1.0.0). Health Data Nexus.


This dataset is composed of merged CT scans of the cervical spine in multi-frame Digital Imaging and Communications in Medicine (DICOM) format, collected from emergency patients at St. Michael's Hospital between May 1, 2004 and July 1, 2021. Images are provided in the axial axis and bone window kernel. The dataset includes a total of 1016 CT scans: 508 with fracture(s) matched with an equal number of CT scans without cervical spine fracture. Imaging studies positive for cervical spine fracture were annotated with a segmented mask over the fracture region. The segmentation provide users with pixel level information regarding fracture features and is provided as an associated DICOM file for each positive scan.


Emergency room visits for trauma account for a significant proportion of total emergency room visits and result in significant health care expenditure. An Ontario Injury Data Report reported 850,000 visits to emergency rooms due to falls and 2.2 million visits for all injury mechanisms combined between April 2014 and March 2016. In a significant portion of emergency room visits for trauma there is sufficient concern of a cervical spine injury to warrant a CT scan of the cervical spine. Patients in whom a cervical spine injury is suspected are typically kept immobilized with a cervical spine collar until a cervical spine injury can be ruled out using a CT scan.

Cervical spine CT scans are typically interpreted by a radiologist, which can be a time consuming process. In addition, many hospitals may not have radiology coverage at all hours, which can contribute to prolonged emergency room stays. While many patients undergo a CT scan of the cervical spine, only a minority (10-20%) will have an acute abnormality on their CT scan that would prompt further management. In the majority of patients, a cervical spine injury can be ruled out using the results of a CT scan in combination with a clinical assessment.

Automated fracture detection within CT scans of the cervical spine could assist radiologists and emergency room physicians in rapidly identifying patients with normal, unconcerning CT scans in whom a cervical spine injury can be ruled out. In addition, such a framework could flag patients with abnormalities for further review.

This dataset was curated for the development and validation of machine learning models which could accurately detect acute fractures on cervical spine CT scans.


Our institutional radiology information system (syngo, Siemens Medical Solutions USA, Inc., Malvern, PA) was searched using Nuance mPower (Nuance Communications, Burlington, MA) to identify emergency patients that underwent a non-contrast CT examination of the cervical spine between May 1, 2004 and July 1, 2021. The associated reports were exported and classified as positive or negative for cervical spine fracture by a board certified radiologist. Images in Digital Imaging and Communications in Medicine (DICOM) format were downloaded from our institutional Picture Archiving and Communications System (Carestream PACS, Carestream Health, Rochester, New York). Custom Python scripts and manual review were utilized to limit DICOM images to axial bone window kernel images of the cervical spine. Images underwent de-identification using RSNA Anonymizer (RSNA, Oak Brook, IL), an open source DICOM de-identification tool developed by the Radiological Society of North America, that complies with Health Insurance Portability and Accountability Act (HIPAA) guidelines. A custom Python script was then used to remove all non-essential DICOM metadata data. A manual review of DICOM images was performed to ensure no private health information was included in the dataset. CT scans were originally stored in a legacy DICOM structure with one DICOM file per slice. For ease of use, we merged together slices into a single multi-frame DICOM file following the Legacy Converted Enhanced Computed Tomography Image Information Object Definition.

Each CT scan was reviewed and annotated independently by three radiologists. Fractures were delineated by drawing regions of interest in ITK-SNAP on each image that contained a fracture. A final ground truth mask was consolidated using a “best 2 out of 3” approach, whereby the resultant mask is the intercept of two masks with the best concordance. Every CT scan positive for fracture has a single DICOM file containing the masks following the Segmentation Information Object Definition.

Data Description

  • Directory dicom: Contains 1016 axial plane CT scans in a multi-frame DICOM format, with 508 negative for cervical spine fracture and 508 positive for cervical spine fracture. Each CT scan is named in the format [Patient ID]_[Accession Number].dcm. A scan is considered positive for fracture if at least one constituent image was found to have a cervical spine fracture.
  • Directory segmentation: Contains 508 merged segmentation masks in a Segmentation DICOM format. The DICOM header for each segmentation provides a structured description of the 14 possible segmentation labels: one for each cervical vertebrae level for acute fractures, and one for each cervical vertebrae level for chronic fractures.
  • File metadata.csv: The Patient_ID and Accession_Number are used as the identifier for each case. The column dicom_path provides the full path name to each case. Additional information provided includes Patient_Age (binned), Patient_Sex and Protocol. For positive scans, the column mask_path will give you the full path to the corresponding segmentation file for each case. Columns c1, c2, c3, c4, c5, c6, and c7 will provide information regarding presence or absence of fracture at each of the 7 cervical vertebrae levels. A value of 0 indicates no fracture, A indicates an acute fracture, C indicates a chronic fracture, and B indicates fractures of both acute and chronic nature. This means that at a given level, it is possible to have more than one fracture, and those fractures could be of either acute or chronic nature.
  • File annotation.csv: For your convenience, this file lists all the images that are positive for fracture. The column exam_identifer will help you identify the case in the format [Patient ID]_[Accession Number]. The slice column will indicate the slice number of interest, cspine_level tells you the cervical spine vertebrae level that the mask represents, while the acute_chronic column indicates whether the fracture is acute or chronic.

Usage Notes

The intended purpose of this data is to build a machine learning model that could detect cervical spine fracture(s) on CT imaging. For more information on the technical aspects of model development, please consult the references listed below. Pydicom is one of several packages available to help with analysis of DICOM images and can be installed in a Jupyter Lab instance of the analysis workspace. For additional information on the format in which data is presented, Legacy Converted Enhanced CT Image IOD can be found here and information on Multi-frame Functional Groups Module can be found here.

The usage.ipynb notebook provides example code for loading multi-frame DICOM images and segmentations into Python for analysis.

Release Notes

v1.0 is the initial release of the dataset.


This project has been approved by Unity Health REB, protocol #21-206.


We would like to thank Blair Jones for his valuable support during the curation of this dataset.

We would also like to thank Zamir Merali for his valuable contributions.

Conflicts of Interest

The authors have no conflicts of interest to declare.


  1. Salehinejad, H., Ho, E., Lin, H. M., Crivellaro, P., Samorodova, O., Arciniegas, M. T., ... & Colak, E. (2021, April). Deep sequential learning for cervical spine fracture detection on computed tomography imaging. In 2021 IEEE 18th International Symposium on Biomedical Imaging (ISBI) (pp. 1911-1914). IEEE.


Access Policy:
Only credentialed users who sign the DUA can access the files.

License (for files):
Health Data Nexus Contributor Review Health Data License 1.0

Data Use Agreement:
T-CAIREM Data Use Agreement

Required training:
TCPS 2: CORE 2022
ISED Cybersecurity Training for Researchers
Health Data Nexus Data User Code of Conduct Training

Corresponding Author
You must be logged in to view the contact information.