Convolutional Neural Network for the Conversion of Gadolinium-Enhanced T1-weighted MRI to Non-Gadolinium T1w Scans

The CNN model outputs are stored at /project/6050199/akhanf/cfmm-bids/data/Lau/degad/snakemake/snakemake_CNN/output

Checkpoint is stored at /project/6050199/akhanf/cfmm-bids/data/Lau/degad/snakemake/snakemake_CNN/output/<model>/checkpoint.pt

Output niftii from a test subject is stored at /project/6050199/akhanf/cfmm-bids/data/Lau/degad/snakemake/snakemake_CNN/output/<model>/test

1. Ensure you are logged onto graham. To clone the MRI-degad repo onto your scratch run: git clone [email protected]:fogunsan/MRI-DeGad.git or git clone https://github.com/fogunsan/MRI-DeGad.git
2. A venv can be created using kslurm with the following command: kpy create venv_degad
3. Following venv creation, cd into the cloned directory and run: pip install -r venv_requirements.txt

16 mm and 32 mm pregenerated patches (outputs from preprocessing pipeline) are stored at /project/6050199/akhanf/cfmm-bids/data/Lau/degad/snakemake/derivatives/patches. Can run snakemake model training workflow to output model checkpoints and output images in output/ directory of snakemake_CNN or snakemake_GAN directory. Can vary arguments at top of Snakefile ensuring max combinations in a single job does not surpass 4 due to time limits. Search space for CNN: Patch size: [15,31] Batch size: [32,64,128] Learning rate: [0.01,0.001,0.005,0.0001] Initial number of filters: [16,32,64] Number of Convolutions: [2,3] Loss: [MAE]

1. cd into cloned repo.
2. cd snakemake_CNN or cd snakemake_GAN, depending on what model you are training
3. mkdir output if it does not already exist
4. Adjust arguments at top of Snakefile. Can check /project/6050199/akhanf/cfmm-bids/data/Lau/shared/training/snakemake_CNN/output to ensure combination has not already been tested
5. In scripts/run_training.sh file, change repo variable to one located on your scratch.
6. bash scripts/run_training.sh

1. Sufficient computational resources must be requested. Sample command: regularInteractive -n 16 -m 64000 -t 1 -g

2. The venv can then be loaded: kpy load venv_train_degad

Checkpoints from trained MRI-degad-CNN models are located at: /project/6050199/akhanf/cfmm-bids/data/Lau/degad/snakemake/snakemake_CNN/output

Script to run inference on a directory of gad images is located at: /project/6050199/akhanf/cfmm-bids/data/Lau/degad/shared/inference/inference_degad_CNN.py

3. To run inference script: cd /project/6050199/akhanf/cfmm-bids/data/Lau/degad/shared/inference python3 inference_degad_CNN.py --checkpoint /project/6050199/akhanf/cfmm-bids/data/Lau/shared/training/snakemake_CNN/output/<model>/checkpoint.pt --gad_direc <input_gad_dir> --output_dir <output_degad_dir> --degad_ds

**Note that directory of gad images must be in non-bids format and that the degad output directory is in bids format ** using the optional --degad_ds flag indicates that bids files (CHANGES, scans.json, dataset_description.json, participants.json, README) from original seeg/dbs degad dataset (/project/6050199/akhanf/cfmm-bids/data/Lau/degad/bids) will be output to make output directory fmriprep ready. If using an external dataset, need to manually input those files afterwards for fMRIPrep to run. **

sample command: python3 inference_degad_CNN.py --checkpoint /project/6050199/akhanf/cfmm-bids/data/Lau/degad/snakemake/snakemake_CNN/output/patch-16_batch-128_LR-0.001_filter-32_depth-3_convs-2_loss-mae/checkpoint.pt --gad_direc ../../derivatives/test_set/ --output_dir ../../derivatives/test_set/test_inference/

Path to degad non-gad and gad bids directory on Graham: /project/6050199/akhanf/cfmm-bids/data/Lau/degad/bids
**Preprocessing code still has my scratch directories, will update to refer to project directory

0-Subject_list.txt
What it does: Holds all 55 subjects that underwent DBS or SEEG at LHSC

1-running fMRIPrep
What it does: Run fMRIPrep on the gad and nongad bids dataset. Outputs have N4 bias correction applied

Terminal command:
bidsBatch -j Long fmriprep_21.0.0 /project/6050199/akhanf/cfmm-bids/data/Lau/degad/derivatives/degad_bids_gad /home/fogunsan/projects/ctb-akhanf/cfmm-bids/Lau/degad/derivatives/fmriprep/gad participant --anat-only --skip_bids_validation --omp-nthreads 8 --nprocs 16

Input directory: /project/6050199/akhanf/cfmm-bids/data/Lau/degad/derivatives/degad_bids_gad and /project/6050199/akhanf/cfmm-bids/data/Lau/degad/derivatives/degad_bids_nongad

Output directory: /project/6050199/akhanf/cfmm-bids/data/Lau/degad/derivatives/fmriprep/gad and /project/6050199/akhanf/cfmm-bids/data/Lau/degad/derivatives/fmriprep/nongad

2-isotropic_resampling.ipynb
What it does: Resamples images to 1 mm volumetric resolution

Input directories: /project/6050199/akhanf/cfmm-bids/data/Lau/degad/derivatives/fmriprep/gad and /project/6050199/akhanf/cfmm-bids/data/Lau/degad/derivatives/fmriprep/nongad

Output directory: /project/6050199/akhanf/cfmm-bids/data/Lau/degad/derivatives/resampled/gad and /project/6050199/akhanf/cfmm-bids/data/Lau/degad/derivatives/resampled/nongad

3-run_greedy.sh
What it does: Performs a rigid and affine transformation, putting nongad images to gad space

Input directories: /project/6050199/akhanf/cfmm-bids/data/Lau/degad/derivatives/resampled/gad and /project/6050199/akhanf/cfmm-bids/data/Lau/degad/derivatives/resampled/nongad

Output directory: /project/6050199/akhanf/cfmm-bids/data/Lau/degad/derivatives/greedy

4-registration_QC.ipynb
What it does: Outputs html file to perform qualitative assessment of the registrations (with preference for rigid registrations) and exclude subjects with pathology

Input directory: /project/6050199/akhanf/cfmm-bids/data/Lau/degad/derivatives/greedy

Output QC file: /project/6050199/akhanf/cfmm-bids/data/Lau/degad/derivatives/registration_QC_0.html

QC Google Sheets: https://docs.google.com/spreadsheets/d/19TnoOD47vY6vuqSzujC0RgwMGdMhtJWCleeeG325nNo/edit?usp=sharing

Passing subject directory: /project/6050199/akhanf/cfmm-bids/data/Lau/degad/derivatives/passing_dataset/

5-run_fcm_whitestripe_norm.sh
What it does: Fuzzy c means normalization normalizes on a subject level the nongad image to the mean of the white matter (using a fMRIPrep WM segmentation mask) which is scaled to 1. Whitestripe normalization performs Z-score normalization based on the intensity values of normal appearing white matter

Input directory: /project/6050199/akhanf/cfmm-bids/data/Lau/degad/derivatives/passing_dataset/

Command to download library: pip install intensity-normalization (https://github.com/jcreinhold/intensity-normalization)

Output directories: /project/6050199/akhanf/cfmm-bids/data/Lau/degad/derivatives/normalized_fcm/ and /project/6050199/akhanf/cfmm-bids/data/Lau/degad/derivatives/normalized_ws/

6a-train_degad_CNN.ipynb
What it does: File to train degad model using a convolutional neural network (U-Net) implementation.

Input directories: /project/6050199/akhanf/cfmm-bids/data/Lau/degad/derivatives/passing_dataset/ and /project/6050199/akhanf/cfmm-bids/data/Lau/degad/derivatives/normalized_fcm/

Output directory: /project/6050199/akhanf/cfmm-bids/data/Lau/degad/derivatives/UNET_outputs/

Command to run:

1. convert to python script: jupyter nbconvert testnotebook.ipynb --to python
2. kbatch gpu -j Long --venv venv_train_degad python 6a-train_degad_CNN.py

6b-train_degad_GAN.ipynb
What it does: File to train degad model using a generative adversarial network (GAN) implementation

Input directory: /project/6050199/akhanf/cfmm-bids/data/Lau/degad/derivatives/passing_dataset/ and /project/6050199/akhanf/cfmm-bids/data/Lau/degad/derivatives/normalized_fcm/

Output directory:/project/6050199/akhanf/cfmm-bids/data/Lau/degad/derivatives/GAN_outputs/

Command to run:

1. convert to python script: jupyter nbconvert testnotebook.ipynb --to python
2. kbatch gpu -j Long --venv venv_train_degad python 6b-train_degad_GAN.py