About Dataset

This dataset contains metadata extracted from training DICOM image files used in the RSNA 2022 Cervical Spine Fracture Detection competition.

Files Included

meta-train: Original extracted metadata (use with caution)
meta-train-clean: Cleaned version of meta-train, easier to use
meta-segmentations: Metadata for images with segmentations, includes correct vertebra labels (C1–C7) extracted from segmentation masks
meta-segmentation-clean: Cleaned version of meta-segmentations
meta-train-with-vertebrae: Metadata for all training images, includes 88% accurate Random Forest predictions of the vertebra in each image
train-segmented: Metadata for all train images with 95% accurate EffNetV2 vertebrae predictions (based on provided notebook)
train-vert-fold4: Version of train-segmented cleaned and enriched via an image+tabular model; includes extra feature-engineered columns
train-vert: Final ensembled predictions from train-segmented and train-vert-fold4