doc/dataset_skm-tea.md

Preparing SKM-TEA Data

MedVision is designed to download raw data from source and the annotations from our Huggingface dataset repo. For access to the raw data of the SKM-TEA dataset, you need to register at https://aimi.stanford.edu/datasets/skm-tea-knee-mri.

This tutorial is about:

• What data to download from the source in order to use the annotations from MedVision?
• How to preprocess the raw data?
• How to structure the preprocessed files?

What data to download?

• Download images from dicoms
• Download masks from segmentation_masks/raw-data-track to Masks

# Files from SKM-TEA source
├── v1-release
    ├── dicoms
        ├── MTR_001.tar.gz
        ├── ...
    ├── segmentation_masks
        ├── raw-data-track
            ├── MTR_001.nii.gz
            ├── ...

# local file structure
├── [your folder]
    ├── Masks
        ├── MTR_001.nii.gz
        ├── ...
    ├── dicoms
        ├── MTR_001.tar.gz
        ├── ...

How to preprocess data?

⚠️ DO NOT change file or folder names. See why here.

1. Go to the parent folder of dicoms, install dcm2niix, add processing script

   # local file structure
   ├── [your folder]
       ├── Masks
       ├── dicoms
       ├── prepare-skm-tea.sh # <-- add this
   

   # prepare-skm-tea.sh
   
   # Step 1: extract files
   mkdir -p dicoms-ext
   for f in dicoms/*.tar.gz; do
       tar -xvzf "$f" -C dicoms-ext
   done
   
   # Step 2: convert dcm to nii.gz
   mkdir -p niigz
   for d in dicoms-ext/MTR_*; do
       dcm2niix -z y -f %f -o niigz/ $d
   done
   
   # Step 3: move files
   mkdir -p Images-E1 Images-E2
   for f in niigz/*_e1.nii.gz; do
       base=$(basename "$f")             # remove path prefix
       mv "$f" "Images-E1/${base/_e1/}"  # rename and move
   done
   for f in niigz/*_e2.nii.gz; do
       base=$(basename "$f")
       mv "$f" "Images-E2/${base/_e2/}"
   done
   
   # clean up
   rm -r niigz
   rm -r dicoms-ext
   

   Expected file structure after running bash prepare-skm-tea.sh

   # local file structure
   ├── [your folder]
       ├── Masks
       ├── dicoms
       ├── Images-E1
       ├── Images-E2
       ├── prepare-skm-tea.sh
   

2. Put these folders into SKM-TEA-nii and upload the compressed file to your HF dataset repo

   # DO NOT use other folder names
   ├── SKM-TEA-nii
       ├── Images-E1
       ├── Images-E2
       ├── Masks
   

   Your HF repo should contain SKM-TEA-nii.zip

3. Set env var whenever you load SKM-TEA data from MedVision

   import os
   os.environ["MedVision_SKMTEA_HF_ID"] = <your-hf-dataset-id>