This program creates a NIFTI file showing the degrees of saliency of voxels in calculating the Loes score of the input anatomical file. The Captum library is used for this.
A baseline approach for computing input attribution. It returns absolute values of the gradients with respect to inputs.
More details about the approach can be found in the following paper: Deep Inside Convolutional Networks: Visualising Image Classification Models and Saliency Maps by Karen Simonyan, Andrea Vedaldi, and Andrew Zisserman.
The program takes three arguments:
- input: the file path of the input NIFTI file
- input: the file path of the model file
- output: the file path of the saliency NIFTI file to be created
Here is an example invocation:
python saliency.py input.nii.gz output.nii.gz loes_scoring_03.pt
There is a virtual environment here that contains all the necessary dependencies:
/home/miran045/reine097/projects/loes-scoring-explainability/venv
The output NIFTI saliency map can be used as an overlay for the original NIFTI file and be viewed in a NIFTI viewer such as FSLeyes.
