RSNA Cervical Spine Fracture Detection - RSNA/AI-Challenge-Data GitHub Wiki
Over 1.5 million spine fractures occur annually in the United States alone resulting in over 17,730 spinal cord injuries annually. The most common site of spine fracture is the cervical spine. There has been a rise in the incidence of spinal fractures in the elderly and in this population, fractures can be more difficult to detect on imaging due to degenerative disease and osteoporosis. Imaging diagnosis of adult spine fractures is now almost exclusively performed with computed tomography (CT). Quickly detecting and determining the location of any vertebral fractures is essential to prevent neurologic deterioration and paralysis after trauma. RSNA has teamed with the American Society of Neuroradiology (ASNR) and the American Society of Spine Radiology (ASSR) to create this ground truth dataset, collecting imaging data from twelve sites on six continents, including approximately 2,000 CT studies. Spine radiology specialists from the ASNR and ASSR provided expert image level annotations these studies to indicate the presence, vertebral level and location of any cervical spine fractures.
Description
The dataset is contained in a Zip archive that includes both DICOM image files (.dcm) and tabular annotation files (.csv). A detailed description of the dataset is provided in Lin HM et al. "The RSNA Cervical Spine Fracture CT Dataset," Radiology: Artificial Intelligence 2023;5:5 (https://doi.org/10.1148/ryai.230034).
License
You may access and use these de-identified imaging datasets and annotations (“the data”) for non-commercial purposes only, including academic research and education, as long as you agree to abide by the following provisions: Not to make any attempt to identify or contact any individual(s) who may be the subjects of the data. If you share or re-distribute the data in any form, include a citation to the “Cervical Spine Fracture Detection Dataset, Copyright RSNA, 2023” as follows: Lin HM, et al. The RSNA Cervical Spine Fracture CT Dataset [10.1148/ryai.230034]. Radiology: Artificial Intelligence 2023;5:5."
Tutorial
Files
train.csv Metadata for the train test set.
- StudyInstanceUID - The study ID. There is one unique study ID for each patient scan.
- patient_overall - One of the target columns. The patient level outcome, i.e. if any of the vertebrae are fractured.
- C[1-7] - The other target columns. Whether the given vertebrae is fractured. See this diagram for the real location of each vertbrae in the spine.
test.csv Metadata for the test set prediction structure. Only the first few rows of the test set are available for download.
- row_id - The row ID. This will match the same column in the sample submission file.
- StudyInstanceUID - The study ID.
- prediction_type - Which one of the eight target columns needs a prediction in this row.
[train/test]_images/[StudyInstanceUID]/[slice_number].dcm The image data, organized with one folder per scan. Expect to see roughly 1,500 scans in the hidden test set. Each image is in the dicom file format. The DICOM image files are ≤ 1 mm slice thickness, axial orientation, and bone kernel. Note that some of the DICOM files are JPEG compressed. You may require additional resources to read the pixel array of these files, such as GDCM and pylibjpeg.
sample_submission.csv A valid sample submission.
- row_id - The row ID. See the test.csv for what prediction needs to be filed in that row.
- fractured - The target column.
train_bounding_boxes.csv Bounding boxes for a subset of the training set.
segmentations/ Pixel level annotations for a subset of the training set. This data is provided in the nifti file format.
A portion of the imaging datasets have been segmented automatically using a 3D UNET model, and radiologists modified and approved the segmentations. The provided segmentation labels have values of 1 to 7 for C1 to C7 (seven cervical vertebrae) and 8 to 19 for T1 to T12 (twelve thoracic vertebrae are located in the center of your upper and middle back), and 0 for everything else. As we focused on the cervical spine, all scans have C1 to C7 labels but not all thoracic labels.
Please be aware that the NIFTI files consist of segmentation in the sagittal plane, while the DICOM files are in the axial plane. Please use the NIFTI header information to determine the appropriate orientation such that the DICOM images and segmentation match. Otherwise, you run the risk of having the segmentations flipped in the Z axis and mirrored in the X axis.