Published under a CC BY 4.0 license. Supplemental material is available for this article. This dataset provides vertebral segmentation masks for spine CT images and annotations of vertebral fractures or abnormalities per vertebral level; it is available from https://osf.io/nqjyw/ This public CT dataset holds 160 image series of 141 patients including segmentation masks of 1725 fully visualized vertebrae; it is split into a training dataset (80 image series, 862 vertebrae), a public validation dataset (40 image series, 434 vertebrae), and a secret test dataset (40 image series, 429 vertebrae, to be released in December 2020). Metadata include annotations of vertebral fractures using the semiquantitative method by Genant and of instances of foreign material per vertebral level, as well as opportunistic measurements of lumbar bone mineral density per patient.

As part of our continuing effort to change the bone health landscape for the better, we have collaborated with University Hospital Brussels and KU Leuven to reinvigorate the current methodology of detecting vertebral (spinal) fractures, because of osteoporosis, which often go undetected. Considered a silent epidemic in bone health; osteoporosis is estimated to affect 200 million people worldwide and is the most common bone disease, resulting in more than 8.9 million fragility fractures each year around the globe. Of these, vertebral fractures are the most common, with one occurring every 22 seconds worldwide in men and women over age 50. The impact of these vertebral fractures can be distressing as they often lead to back pain, loss of height, deformity, immobility, increased number of bed days, and even reduced pulmonary function. The impact on a patient's mental health is also significant with reports of loss of self-esteem, distorted body image and depression.

Since our task is detection and not segmentation, correctly predicting only a sufficient amount of voxels around the vertebra centroid is needed to detect normal or fractured vertebrae in an image. We leverage this observation to construct 3D label images for our training database in a semi-automated fashion. First, radiologist S.R. created a text file with annotations for every vertebra present in the field of view as described in section 2. Next, J.N. enriched these labels with 3D centroid coordinates by manually localizing every vertebra centroid in the image using MeVisLab [8]. This step required an average of less than two minutes per image in our dataset. Finally, we extended the method described by Glocker et al. [6] to automatically generate 3D label images from these sparse annotations. The resulting label images contain ellipsoids (flattened along the longitudinal axis for fractured vertebrae) around each vertebra centroid annotated with the ground truth class label provided by the radiologist (combining mild, moderate and severe fractures into one fracture class because of the low number of examples per class, see Figure 1).