Skip to content

interpret - Interpretation with gradient maps

This functionality allows interpreting pretrained models by computing a mean saliency map across a group of images. It takes as input MAPS-like model folders.

Prerequisites

Please check which preprocessing needs to be performed in the maps.json file of the MAPS. If it has not been performed, execute the preprocessing pipeline as well as clinicadl extract to obtain the tensor versions of the images.

Running the task

This task can be run with the following command line:

clinicadl interpret INPUT_MAPS_DIRECTORY DATA_GROUP NAME
where:

  • INPUT_MAPS_DIRECTORY (Path) is a path to the MAPS folder containing the model which will be interpreted.
  • DATA_GROUP (str) is a prefix to name the files resulting from the interpretation task.
  • NAME (str) is the name of the saliency map task.

data group consistency

For ClinicaDL, a data group is linked to a list of participants / sessions and a CAPS directory. When performing a prediction, interpretation or tensor serialization the user must give a data group. If this data group does not exist, the user MUST give a caps_path and a tsv_path. If this data group already exists, the user MUST not give any caps_path or tsv_path, or set overwrite to True.

Optional arguments:

  • Computational resources
    • --gpu / --no-gpu (bool) Uses GPU acceleration or not. Default behaviour is to try to use a GPU. If not available an error is raised. Use the option --no-gpu if running in CPU.
    • --n_proc (int) is the number of workers used by the DataLoader. Default: 2.
    • --batch_size (int) is the size of the batch used in the DataLoader. Default: 2.
  • Model selection
    • --selection_metrics (List[str]) is a list of metrics to find the best models to evaluate. Default will predict the results for best model based on the loss only.
  • Data management
    • --participants_tsv (Path) is a path to a directory containing one TSV file per diagnosis (see output tree of getlabels). Default will use the same participants as those used during the training task.
    • --caps_directory (Path) is the path to a CAPS hierarchy. Default will use the same CAPS as during the training task.
    • --multi_cohort (bool) is a flag indicated that multi-cohort classification is performed. In this case, caps_directory and tsv_path must be paths to TSV files.
    • --diagnoses (List[str]) if tsv_file is a split directory, then will only load the labels wanted. Default will look for the same labels used during the training task.
  • Other options
    • --target_node (int) is the node the gradients explain. By default, it will target the first output node.
    • --save_individual (bool) is an option to save individual saliency maps in addition to the mean saliency map.

Outputs

Results for the DATA_GROUP level are stored in the results folder given by INPUT_MAPS_DIRECTORY, according to the following file system:

<maps_directory>
    ├── fold-0  
    ├── ...  
    └── fold-<fold>
        └── best-<metric>
                └── <data_group>
                    └── interpret-<name>
                        ├── mean_<mode>-<k>_map.pt
                        └── sub-<i>_ses-<j>_<mode>-<k>.pt

  • mean_<mode>-<k>_map.pt is the tensor of the mean saliency map for mode k across the data set used (always saved),
  • sub-<i>_ses-<j>_<mode>-<k>.pt is the tensor of the saliency map for participant i, session j and mode_id k (saved only if flag --save_individual was given).