meqili/Deep-3D-CNNs-for-MRI-Classification-with-Alzheimer-s-Disease-And-Grad-CAM-Visualization

The pre-processing codes for 3D MRI data are also provided stp by step in my github, please check this link if you need.

This repo has the receipts for buiding singularity containers, or you can setup the environment by yourself based on the receipts. Please check the folder singularity_receipt.

Citation

If you are using this repository, please cite this article Li Q, Yang MQ. 2021. Comparison of machine learning approaches for enhancing Alzheimer’s disease classification. PeerJ 9:e10549 https://doi.org/10.7717/peerj.10549

Singularity or Requirment

#buiding singularity containers
sudo singularity build resnet_cnn_mri.def classify.img
sudo singularity build cam_vis.img cam_vis.def

1. MRI Classification

1.1 Data Preparation (folder data)

File Name	Description
all_metadata.csv	All image data
metadata.csv	Training dataset
test.csv	Test dataset

1.2 Classification

module load singularity

##Use the VGG network:
nohup singularity exec --nv classify.img python3.5 scripts/vgg_3d_pred.py > vgg.out &

##Use the Resnet network:
nohup singularity exec --nv classify.img python3.5 scripts/res_3d_pred.py >resnet.out &

1.3 Result

Please check the folders results_vgg and results_resnet for more details.

2. Grad-CAM Visualization

1.1 Grad-CAM

After the classification, the models have learned the weights from images. Choose one image in which you would love to view the discriminative regions. In this example, I used S117504-reg.nii.gz. And also, you can change to the layer that you want to visualize.

Usage:

##Without Singularity
python3.5  python_script  imgpath prefix
##With Singularity
Singularity  exec --nv classify.img  python3.5 python_script  imgpath  prefix

For example:

qli@gpu001$ singularity exec --nv classify.img python3.5 scripts/vgg_3d_grad_cam.py /home/qli/AlzheimerClassify/5.Resize/S117504-reg.nii.gz S117504
Using gpu device 0: Tesla P100-PCIE-12GB (CNMeM is disabled, cuDNN 5110)
[[ 0.00110552  0.99889451]]
test_loss 0.001106127048842609
-----------
~/mri_classif


qli@gpu001$ singularity exec --nv classify.img python3.5 scripts/resnet_3d_grad_cam.py /home/qli/AlzheimerClassify/5.Resize/S117504-reg.nii.gz S117504
Using gpu device 0: Tesla P100-PCIE-12GB (CNMeM is disabled, cuDNN 5110)
[[ 0.12858798  0.87141204]]
test_loss 2.0511419773101807
-----------
~/mri_classif

From the examples, the scripts will return some info below:

[[ 0.12858798  0.87141204]]
test_loss 2.0511419773101807

The two number inside the double square brackets stands for the possibilities of this image is classified as normal and Alzheimer’s respectively.

1.2 Debugging

If you got a problem like below:

_tkinter.TclError: no display name and no $DISPLAY environment variable

Here is the solution:

echo "backend: Agg" > ~/.config/matplotlib/matplotlibrc

1.3 Visualization

After running the Grad-CAM, two .npz files (one contains the discriminative info and another one contains the MRI data) will be saved in folder npz_res or npz_vgg. Scripts in this section do not need to run on the GPU cards.

#Usage:
singularity exec cam_vis.img python3.6 scripts/visualize_cam.py cam_npz/file/path mri_npz/file/path prefix

#Examples:

singularity exec cam_vis.img python3.6 scripts/visualize_cam.py npz_vgg/S117504_cam_conv4c.npz npz_vgg/S117504_mri.npz S117504_vgg

singularity exec cam_vis.img python3.6 scripts/visualize_cam.py npz_res/S117504_cam_voxres9_conv2.npz npz_res/S117504_mri.npz S117504_res

meqili / Deep-3D-CNNs-for-MRI-Classification-with-Alzheimer-s-Disease-And-Grad-CAM-Visualization