DEEP QC

Code for the paper Vladimir S. Fonov, Mahsa Dadar, The PREVENT-AD Research Group, D. Louis Collins "DARQ: Deep learning of quality control for stereotaxic registration of human brain MRI" https://doi.org/10.1016/j.neuroimage.2022.119266

Updated version of the previosly available "Deep learning of quality control for stereotaxic registration of human brain MRI"

Installation (Python version) using conda for inference

• CPU version

  conda install pytorch-cpu==1.7.1 torchvision==0.8.2 cpuonly -c pytorch 
  conda install scikit-image
  

• GPU version

  conda install pytorch==1.7.1 torchvision==0.8.2 cudatoolkit=<your cuda version>  -c pytorch 
  conda install scikit-image
  

• (optional) minc toolkit and minc2-simple

  conda install -c vfonov minc-toolkit-v2 minc2-simple
  

• (optional) nibabel (for analyzing Freesurfer output)

  pip install nibabel
  

Running

Inference

• Inference using jpeg files generated by minc_aqc.pl script: python3 python/aqc_apply.py --net <network> --image <image base>
• Inference using minc files in stereotaxic space: python3 python/aqc_apply.py --net <network> --volume <input.mnc>
• Inference using Freesurfer output: python3 python/aqc_apply.py --net <network> --freesurfer <freesurfer subject directory>

Training

• Training in python directory run_all_experiments.sh - will try to train all networks

Dependencies

• trainig dependencies: scikit-image tensorboard,
• for inference directly on minc files minc2-simple
• for inference on freeserfer files: nibabel
• minc2-simple (optional): https://github.com/vfonov/minc2-simple

Files

• Shell scripts:
  • download_minimal_models.sh - download QCResNET-18 with reference pretrained model to run automatic qc (43mb)
  • download_all_models.sh - download all pretrained models to run automatic qc
• Directory python:
  • run_all_experiments.sh - run experiments with different versions of ResNet and SquezeNet
  • aqc_apply.py - apply pre-trained network
  • aqc_convert_to_cpu.py- helper script to convert network from GPU to CPU
  • aqc_data.py - module to load QC data
  • aqc_training.py - deep nearal net training script
  • model/resnet_qc.py - module with ResNET implementation, based on https://github.com/pytorch/vision/blob/master/torchvision/models/resnet.py
  • model/util.py - various helper functions
  • *.R - R scripts to generete figures for the paper
• Image files:
  • mni_icbm152_t1_tal_nlin_sym_09c_0.jpg,mni_icbm152_t1_tal_nlin_sym_09c_1.jpg,mni_icbm152_t1_tal_nlin_sym_09c_2.jpg - reference slices, needed for both training and running pretrained model
• results - figures for the paper
• data - reference images

Validating correct operation (requires minc-toolkit and minc2_simple python module)

# create a file with 30 degree rotation transform
param2xfm -rotations 30 0 0  rot_30.xfm
# apply to a template:
itk_resample /opt/minc/share/icbm152_model_09c/mni_icbm152_t1_tal_nlin_sym_09c.mnc --transform rot_30.xfm bad.mnc

# run QC script on good scan
# should print "Pass"
python python/aqc_apply.py --net r18 --volume /opt/minc/share/icbm152_model_09c/mni_icbm152_t1_tal_nlin_sym_09c.mnc

# now on "bad"
# should print "Fail"
python python/aqc_apply.py --net r18 --volume bad.mnc

Validating correct operation on Freesurfer, requires nibabel

# run recon-all on ernie subject from freesurfer installation
recon-all -s ernie -i $SUBJECTS_DIR/sample-001.mgz -i $SUBJECTS_DIR/sample-002.mgz -autorecon1

# run automated QC on freesurfer output
# should pring "Pass"
python python/aqc_apply.py --net r18 --freesurfer $FREESURFER_HOME/subjects/ernie