bsaghafi / 3dcnn-dmri

Keras implementation of a 3DCNN that predicts level of head impact exposure from changes in diffusion MRI

3DCNN-dMRI

Keras Implementation of a 3DCNN that predicts level of head impact exposure from changes in FA images post-pre. This repository includes training, validation, permutation testing and occlusion map for discovering feature importance.

Please cite the following paper if you use the code:

• Quantifying the Association between White Matter Integrity Changes and Subconcussive Head Impact Exposure

Training and Testing

For training and testing the model, run:

3DCNN_diffusion_train_test.py

For testing the pretrained model, run:

3DCNN_diffusion_test_pretrained.py

For performing model selection using 5-fold cross validation, run:

3DCNN_diffusion_ModelSelection_validation.py

In each validation test, the learning rate was reduced by 50% if the ROC AUC did not improve for 10 epochs. An early stopping schedule with look ahead was developed. Training stoped when the network showed no improvement in AUC score for 50 epochs.

Permutation Testing

For performing permutation testing and evaluating the statistical significance of the overall model, run:

permutation_testing.py

Occlusion Maps

For computing occlusion maps for the high impact exposure sample, run:

occlusion_map.py