---
license: gpl-3.0
---
# Dataset structure
- checkpoints: Contains nnunet checkpoint files for hip segmentation (proximal femur), knee segmentation (distal femur and proximal tibia) and ankle segmentation (distal tibia and distal fibula). Three checkpoint files per segmentation model, one for each augmentation scheme (baseline aka no augmentation, default aka nnunet augmentation and mr aka MRI-specific augmentation).
- testdata_raw: Contains the test dataset in raw form, i.e. original dicom series. Per participant (n=20), there are five dicom files: one for the reference exam and four each for the motion-pattern exams (gluteal contraction in high (glutes_hf) and low (glutes_lf) frequency and plantar-/dorsiflexion in high (feet_hf) and low (feet_lf) frequency).
- torsion_format: Contains test data structured for torsional alignment quantification post-processing. Data is organised into reference (no artefact) images, images with mild, moderate and severe artefacts. Each series directory (named participant_[exam], e.g. participant_1_feet_lf is the series resulting from the low-frequency plantar-/dorsiflexion exam for participant #1) contains: 
	- in nifti format: raw images for hip, knee and ankle
	- in nifti format: segmentation masks for hip, knee and ankle (Y_seg.nii.gz = segmentation by MRI-augmented model, Y_seg_baseline = segmentation by baseline model, Y_seg_default = segmentation by default-augmented model)
	- in nifti format: segmentation masks + reference lines for hip, knee, and ankle (Y_ref_)
	- in json format: computed torsion angles. hip_left/right = angle of proximal reference line. knee_femur_left/right = angle of distal femur reference line. femur_left/right = femoral torsion angle. knee_tibia_left/right = angle of proximal tibia reference line. ankle_left/right = angle of distal tibia/fibula reference line. tibia_left/right = tibial torsion angle. If a key is missing, the algorithm failed to compute the corresponding referenc line / angle.
	
# Code
Code is available at github.com/swestfechtel/paper-augmentation