Skip to content

Classification based on machine learning using scikit-learn

Clinica provides a modular way to perform classification based on machine learning. To build its own classification pipeline, the user can combine three modules based on scikit-learn [Pedregosa et al., 2011]:

  • Input (e.g. gray matter maps obtained from T1-weighted MR images, FDG PET images)
  • Algorithm (e.g. support vector machine, logistic regression, random forest)
  • Validation (e.g. K-fold cross validation, repeated K-fold cross validation, repeated hold-out validation)

Prerequisites

You need to have performed the t1-volume pipeline on your T1-weighted MR images and/or the pet-volume pipeline on your PET images.

Dependencies

If you installed the core of Clinica, this pipeline needs no further dependencies.

Classification modules

Input

Two classes corresponding to the voxel-based and the region-based approaches are implemented in input.py:

  • CAPSVoxelBasedInput: all the voxels of the image are used as features.
  • CAPSRegionBasedInput: a list of values stored in a TSV file is used as features. This list corresponds to PET or T1 image intensities averaged over a set of regions obtained from a brain parcellation when running the t1-volume and/or pet-volume pipeline.

Note

The atlases that can be used for the region-based approaches are listed here.

Algorithm

Three classes corresponding to the machine learning-based classification algorithms are implemented in algorithm.py:

  • DualSVMAlgorithm: support vector machine (SVM) algorithm (input: all the data available or a kernel that can be pre-computed)
  • LogisticReg: logistic regression algorithm (input: all the data available)
  • RandomForest: random forest algorithm (input: all the data available)

Each algorithm implements a grid search approach to choose the best parameters for the classification by looking at the value of the balanced accuracy. The area under the receiver operating characteristic (ROC) curve (AUC) is also reported. The labels are automatically assigned based on the diagnoses_tsv file.

Validation

Three classes corresponding to the validation strategies are implemented in validation.py:

  • KFoldCV: K-fold cross validation
  • RepeatedKFoldCV: repeated K-fold cross validation
  • RepeatedHoldOut: repeated hold-out validation

The input is the name of the classification algorithm used.

Running your pipeline

No matter the combination of modules chosen, the inputs necessary are:

  • caps_directory: the folder containing the results of the t1-volume and/or the pet-volume pipeline (where TSV files are stored)
  • subjects_visits_tsv: the TSV file containing the participant_id and the session_id columns
  • diagnoses_tsv: a TSV file where the diagnosis for each participant (identified by a participant ID) is reported (e.g. AD, CN). It allows the algorithm to perform the dual classification (between the two labels reported). Example of a diagnosis TSV file: 
    participant_id    diagnosis
sub-CLNC0001      AD
sub-CLNC0002      CN
sub-CLNC0003      AD
sub-CLNC0004      AD
sub-CLNC0005      CN
  • group_label: the label of the group of subjects studied
  • image_type: a value to set the modality studied (T1w or PET)
  • output_dir: the directory where outputs are saved
  • atlas: the name of the atlas used for the brain parcellation in case of a region-based approach
  • fwhm: the FWHM value in mm used in the t1-volume or pet-volume pipeline
  • modulated: a flag to indicate if, when running the t1-volume pipeline, the image has been modulated or not (on, off)
  • acq_label: label given to the PET acquisition, specifying the tracer used (acq-<acq_label>)
  • suvr_reference_region: reference region used to perform intensity normalization (i.e. dividing each voxel of the image by the average uptake in this region) resulting in a standardized uptake value ratio (SUVR) map. It can be cerebellumPons (used for amyloid tracers) or pons (used for FDG).
  • use_pvc_data: use PET data with partial value correction (True/False). By default, PET data with no PVC are used)
  • precomputed_kernel: to load the precomputed kernel if it exists
  • mask_zeros: a flag to indicate if zero-valued voxels should be taken into account for the classification (True/False)
  • n_iterations: number of times a task is repeated
  • grid_search_folds: number of folds to use for the hyper-parameter grid search (e.g. 10)
  • c_range: range used to select the best value for the C parameter, in the logspace
  • n_threads: number of threads used if run in parallel
  • test_size: percentage (between 0 and 1) representing the size of the test set for each shuffle split
  • balanced: option to balance the weights according to the number of samples
  • penalty: type of penalty ("l2" or "l1")

Tip

Usage examples are available in ml_workflows.py.

Output

Results are saved in the output folder following this hierarchy: 

└── <image-type>
    ├── region_based
    |    └── atlas-<atlas-id>
    |        └── <machine-learning-algorithm>
    |             └── <task1>_vs_<task2>
    |                 ├── classifier
    |                 |    └── iteration-<iteration-number>
    |                 |        ├── mean_results.tsv
    |                 |        ├── results.tsv
    |                 |        └── subjects.tsv
    |                 ├── best_parameters.json
    |                 ├── dual_coefficients.txt
    |                 ├── intersect.txt
    |                 ├── support_vector_indices.json
    |                 ├── weights.nii.gz
    |                 └── weights.txt
    └── voxel_based
        └── smoothing-<fwhm>
            └── <machine-learning-algorithm>
                └── <task1>_vs_<task2>
                    ├── classifier
                    |    └── iteration-<number-iteration>
                    |        ├── mean_results.tsv
                    |        ├── results.tsv
                    |        └── subjects.tsv
                    ├── best_parameters.json
                    ├── dual_coefficients.txt
                    ├── intersect.txt
                    ├── support_vector_indices.json
                    ├── weights.nii.gz
                    └── weights.txt

If image_type is PET: 

└── <image-type>
    └── region_based/voxel_base
        └── pvc-<pvc>
            └── ...

Describing this pipeline in your paper

Example of paragraph:

These results have been obtained using the machine learning-based classification modules of Clinica [Routier et al; Samper et al., 2018]. Clinica provides a modular way to perform classification based on machine learning by combining different inputs (e.g. gray matter maps obtained from T1-weighted MR images, FDG PET images), algorithms (e.g. support vector machine, logistic regression, random forest) and validation strategies (e.g. K-fold cross validation, repeated K-fold cross validation, repeated hold-out validation). These modules rely on scikit-learn [Pedregosa et al., 2011].

Support