davidshq/miqa-ml

A standalone version of the MIQA ML code.

MIQA Pixel-based Classifier

DON'T Use Visual Studio Code DevContainers!

Normally I love VSC DevContainers but installing the requirements into a devcontainer seems to lock things up. It's a pain.

You can try, maybe your computer is faster, better, smarter than mine (likely).

Setup

• Pull down the repo
• Create a python virtual environment: python3 -m venv venv
• Activate the virtual environment: source venv/bin/activate
• Install required packages: pip install -r ./miqa/learning/requirements.txt

Pre-trained Model Files

This repo includes pre-trained models in the models subdirectory so that you can begin using the neural net without waiting for training to complete.

Run training

Get the data

For example, copy PredictHD_small from the shared Google Drive, as well as files T1_fold0.csv, T1_fold1.csv and T1_fold2.csv from Learning.

Now edit the paths inside T1_fold*.csv files to match the path on your file system, e.g.:

sed -i 's+P:/PREDICTHD_BIDS_DEFACE/+/home/exampleUser/nn_work_dir/PredictHD_small/+g' T1_fold0.csv

Limited 3-Fold Cross Validation

To run training on the 3-fold cross-validation using fold 0 as validation set and folds 1 and 2 as training set, use:

python ./miqa/learning/nn_training.py -f ./T1_fold -c 3 -v 0

Full 3-Fold Cross Validation

This will produce a file called miqa01-val0.pth.

To run full 3-fold cross validation, execute:

python ./miqa/learning/nn_training.py -f ./T1_fold -c 3 --all

This will produce miqa01-val0.pth, miqa01-val1.pth and miqa01-val2.pth.

Run inference

Inference on a Single File

To run inference on a single file, execute:

python ./nn_training.py -m ./models/miqaT1-val0.pth -1 ./images/coronacases_001.nii.gz

The above command uses the neural network weights produced by training on folds 1 and 2.

Inference on One of the Folds

To run inference on one of the folds, e.g. fold 1, use:

python ./miqa/learning/nn_training.py -f ./T1_fold -c 3 -v 1 --evaluate