Merge pull request #673 from gagnonanthony/pcadwi

Adding PCA script for dimension reduction of metrics

scilpy/io/fetcher.py

@@ -102,7 +102,7 @@ def get_testing_files_dict():
              '3e27625a1e7f2484b7fa5028c95324cc'],
             'stats.zip':
             ['1vsM7xuU0jF5fL5PIgN6stAH7oO683tw0',
-             'bcc21835cf0bf7210bdc99ba5d8df44b'],
+             'fdad170bcc9d19a2e53b0f29fd793ad4'],
             'anatomical_filtering.zip':
             ['1Li8DdySnMnO9Gich4pilhXisjkjz1-Dy',
              '6f0eff5154ff0973a3dc26db00e383ea'],

scripts/scil_compute_pca.py

@@ -0,0 +1,313 @@
+#!/usr/bin/env python
+# -*- coding: utf-8 -*-
+
+"""
+Script to compute PCA analysis on diffusion metrics. Output returned is all significant principal components
+(e.g. presenting eigenvalues > 1) in a connectivity matrix format. This script can take into account all
+edges from every subject in a population or only non-zero edges across all subjects.
+
+The script can take directly as input a connectoflow output folder. Simply use the --input_connectoflow flag.
+For other type of folder input, the script expects a single folder containing all matrices for all subjects.
+Example:
+        [in_folder]
+        |--- sub-01_ad.npy
+        |--- sub-01_md.npy
+        |--- sub-02_ad.npy
+        |--- sub-02_md.npy
+        |--- ...
+
+The plots, tables and principal components matrices will be outputted in the designated folder from the
+<out_folder> argument. If you want to move back your principal components matrices in your connectoflow
+output, you can use a similar bash command for all principal components:
+for sub in `cat list_id.txt`; do cp out_folder/${sub}_PC1.npy connectoflow_output/$sub/Compute_Connectivity/; done
+
+Interpretation of resulting principal components can be done by evaluating the loadings values for each metrics.
+A value near 0 means that this metric doesn't contribute to this specific component whereas high positive or
+negative values mean a larger contribution. Components can then be labeled based on which metric contributes
+the highest. For example, a principal component showing a high loading for afd_fixel and near 0 loading for all
+other metrics can be interpreted as axonal density (see Gagnon et al. 2022 for this specific example or
+ref [3] for an introduction to PCA).
+
+EXAMPLE USAGE:
+scil_compute_pca.py input_folder/ output_folder/ --metrics ad fa md rd [...] --list_ids list_ids.txt
+"""
+
+# Import required libraries.
+import argparse
+import logging
+
+import matplotlib.pyplot as plt
+import numpy as np
+import pandas as pd
+from sklearn.preprocessing import StandardScaler
+from sklearn.decomposition import PCA
+
+from scilpy.io.utils import (load_matrix_in_any_format,
+                             save_matrix_in_any_format,
+                             add_verbose_arg,
+                             add_overwrite_arg,
+                             assert_output_dirs_exist_and_empty)
+
+
+EPILOG = """
+[1] Chamberland M, Raven EP, Genc S, Duffy K, Descoteaux M, Parker GD, Tax CMW, Jones DK. Dimensionality 
+    reduction of diffusion MRI measures for improved tractometry of the human brain. Neuroimage. 2019 Oct 
+    15;200:89-100. doi: 10.1016/j.neuroimage.2019.06.020. Epub 2019 Jun 20. PMID: 31228638; PMCID: PMC6711466.
+[2] Gagnon A., Grenier G., Bocti C., Gillet V., Lepage J.-F., Baccarelli A. A., Posner J., Descoteaux M., 
+    & Takser L. (2022). White matter microstructural variability linked to differential attentional skills 
+    and impulsive behavior in a pediatric population. Cerebral Cortex. https://doi.org/10.1093/cercor/bhac180
+[3] https://towardsdatascience.com/what-are-pca-loadings-and-biplots-9a7897f2e559
+    """
+
+
+# Build argument parser.
+def _build_arg_parser():
+    p = argparse.ArgumentParser(
+        description=__doc__,
+        formatter_class=argparse.RawTextHelpFormatter,
+        epilog=EPILOG)
+
+    p.add_argument('in_folder',
+                   help='Path to the input folder.')
+    p.add_argument('out_folder',
+                   help='Path to the output folder to export graphs, tables and principal components matrices.')
+    p.add_argument('--metrics', nargs='+', required=True,
+                   help='Suffixes of all metrics to include in PCA analysis (ex: ad md fa rd). They must be '
+                        'immediately followed by the .npy extension.')
+    p.add_argument('--list_ids', required=True, metavar='FILE',
+                   help='Path to a .txt file containing a list of all ids.')
+    p.add_argument('--not_only_common', action='store_true',
+                   help='If true, will include all edges from all subjects and not only common edges (Not recommended)')
+    p.add_argument('--input_connectoflow', action='store_true',
+                   help='If true, script will assume the input folder is a Connectoflow output.')
+
+    add_verbose_arg(p)
+    add_overwrite_arg(p)
+
+    return p
+
+
+def generate_pca_input(dictionary):
+    """
+    Function to create PCA input from matrix.
+    :param dictionary:       Dictionary with metrics as keys containing list of matrices sorted by ids.
+    :return:                 Numpy array.
+    """
+    for metric in dictionary.keys():
+        mat = np.rollaxis(np.array(dictionary[metric]), axis=1, start=3)
+        matrix_shape = mat.shape[1:3]
+        n_group = mat.shape[0]
+        mat = mat.reshape((np.prod(matrix_shape) * n_group, 1))
+        dictionary[metric] = mat
+
+    return np.hstack([dictionary[i] for i in dictionary.keys()])
+
+
+def restore_zero_values(orig, new):
+    """
+    Function to restore 0 values in a numpy array.
+    :param orig:    Original numpy array containing 0 values to restore.
+    :param new:     Array in which 0 values need to be restored.
+    :return:        Numpy array with data from the new array but zeros from the original array.
+    """
+    mask = np.copy(orig)
+    mask[mask != 0] = 1
+
+    return np.multiply(new, mask)
+
+
+def autolabel(rects, axs):
+    """
+    Attach a text label above each bar displaying its height (or bar value).
+    :param rects:   Graphical object.
+    :param axs:     Axe number.
+    :return:
+    """
+    for rect in rects:
+        height = rect.get_height()
+        if height > 0:
+            axs.text(rect.get_x() + rect.get_width() / 2., (height * 1.05),
+                     '%.3f' % float(height), ha='center', va='bottom')
+        else:
+            axs.text(rect.get_x() + rect.get_width()/2., (height*1.05) - 0.15,
+                     '%.3f' % float(height), ha='center', va='bottom')
+
+
+def extracting_common_cnx(dictionary, ind):
+    """
+    Function to create a binary mask representing common connections across the population
+    containing non-zero values.
+    :param dictionary:       Dictionary with metrics as keys containing list of matrices sorted by ids.
+    :param ind:              Indice of the key to use to generate the binary mask from the dictionary.
+    :return:                 Binary mask.
+    """
+    keys = list(dictionary.keys())
+    met = np.copy(dictionary[keys[ind]])
+
+    # Replacing all non-zero values by 1.
+    for i in range(0, len(met)):
+        met[i][met[i] != 0] = 1
+
+    # Adding all matrices together.
+    orig = np.copy(met[0])
+    mask = [orig]
+    for i in range(1, len(met)):
+        orig = np.add(orig, met[i])
+        mask.append(orig)
+
+    # Converting resulting values to 0 and 1.
+    mask_f = mask[(len(met) - 1)]
+    mask_f[mask_f != len(met)] = 0
+    mask_f[mask_f == len(met)] = 1
+
+    return mask_f
+
+
+def apply_binary_mask(dictionary, mask):
+    """
+    Function to apply a binary mask to all matrices contained in a dictionary.
+    :param dictionary:       Dictionary with metrics as keys containing list of matrices sorted by ids.
+    :param mask:             Binary mask.
+    :return:                 Dictionary with the same shape as input.
+    """
+    for a in dictionary.keys():
+        for i in range(0, len(dictionary[a])):
+            dictionary[a][i] = np.multiply(dictionary[a][i], mask)
+
+    return dictionary
+
+
+def main():
+    parser = _build_arg_parser()
+    args = parser.parse_args()
+
+    if args.verbose:
+        logging.getLogger().setLevel(logging.INFO)
+
+    assert_output_dirs_exist_and_empty(parser, args, args.out_folder, create_dir=True)
+
+    subjects = open(args.list_ids).read().split()
+
+    if args.input_connectoflow:
+        # Loading all matrix.
+        logging.info('Loading all matrices from a Connectoflow output...')
+        dictionary = {m: [load_matrix_in_any_format(f'{args.in_folder}/{a}/Compute_Connectivity/{m}.npy')
+                          for a in subjects]
+                      for m in args.metrics}
+    else:
+        logging.info('Loading all matrices...')
+        dictionary = {m: [load_matrix_in_any_format(f'{args.in_folder}/{a}_{m}.npy') for a in subjects]
+                      for m in args.metrics}
+        # Assert that all metrics have the same number of subjects.
+        nb_sub = [len(dictionary[m]) for m in args.metrics]
+        assert all(x == len(subjects) for x in nb_sub), "Error, the number of matrices for each metric doesn't match" \
+                                                        "the number of subject in the id list." \
+                                                        "Please validate input folder."
+
+    # Setting individual matrix shape.
+    mat_shape = dictionary[args.metrics[0]][0].shape
+
+    if args.not_only_common:
+        # Creating input structure using all edges from all subjects.
+        logging.info('Creating PCA input structure with all edges...')
+        df = generate_pca_input(dictionary)
+    else:
+        m1 = extracting_common_cnx(dictionary, 0)
+        m2 = extracting_common_cnx(dictionary, 1)
+
+        if m1.shape != mat_shape:
+            parser.error("Extracted binary mask doesn't match the shape of individual input matrix.")
+
+        if np.sum(m1) != np.sum(m2):
+            parser.error("Different binary mask have been generated from 2 different metrics, \n "
+                         "please validate input data.")
+        else:
+            logging.info('Data shows {} common connections across the population.'.format(np.sum(m1)))
+
+        dictionary = apply_binary_mask(dictionary, m1)
+
+        # Creating input structure.
+        logging.info('Creating PCA input structure with common edges...')
+        df = generate_pca_input(dictionary)
+
+    # Setting 0 values to nan.
+    df[df == 0] = 'nan'
+
+    # Standardized the data.
+    logging.info('Standardizing data...')
+    x = StandardScaler().fit_transform(df)
+    df_pca = x[~np.isnan(x).any(axis=1)]
+    x = np.nan_to_num(x, nan=0.)
+
+    # Perform the PCA.
+    logging.info('Performing PCA...')
+    pca = PCA(n_components=len(args.metrics))
+    principalcomponents = pca.fit_transform(df_pca)
+    principaldf = pd.DataFrame(data=principalcomponents, columns=[f'PC{i}' for i in range(1, len(args.metrics) + 1)])
+
+    # Plot the eigenvalues.
+    logging.info('Plotting results...')
+    eigenvalues = pca.explained_variance_
+    pos = list(range(1, len(args.metrics)+1))
+    plt.clf()
+    plt.cla()
+    fig = plt.figure()
+    ax = fig.add_subplot(1, 1, 1)
+    bar_eig = ax.bar(pos, eigenvalues, align='center', tick_label=principaldf.columns)
+    ax.set_xlabel('Principal Components', fontsize=10)
+    ax.set_ylabel('Eigenvalues', fontsize=10)
+    ax.set_title('Eigenvalues for each principal components.', fontsize=10)
+    ax.margins(0.1)
+    autolabel(bar_eig, ax)
+    plt.savefig(f'{args.out_folder}/eigenvalues.pdf')
+
+    # Plot the explained variance.
+    explained_var = pca.explained_variance_ratio_
+    plt.clf()
+    plt.cla()
+    fig = plt.figure()
+    ax = fig.add_subplot(1, 1, 1)
+    bar_var = ax.bar(pos, explained_var, align='center', tick_label=principaldf.columns)
+    ax.set_xlabel('Principal Components', fontsize=10)
+    ax.set_ylabel('Explained variance', fontsize=10)
+    ax.set_title('Amount of explained variance for all principal components.', fontsize=10)
+    ax.margins(0.1)
+    autolabel(bar_var, ax)
+    plt.savefig(f'{args.out_folder}/explained_variance.pdf')
+
+    # Plot the contribution of each measures to principal component.
+    component = pca.components_
+    output_component = pd.DataFrame(component, index=principaldf.columns, columns=args.metrics)
+    output_component.to_excel(f'{args.out_folder}/loadings.xlsx', index=True, header=True)
+    plt.clf()
+    plt.cla()
+    fig, axs = plt.subplots(2)
+    fig.suptitle('Graph of the contribution of each measures to the first and second principal component.', fontsize=10)
+    bar_pc1 = axs[0].bar(pos, component[0], align='center', tick_label=args.metrics)
+    bar_pc2 = axs[1].bar(pos, component[1], align='center', tick_label=args.metrics)
+    axs[0].margins(0.2)
+    axs[1].margins(0.2)
+    autolabel(bar_pc1, axs[0])
+    autolabel(bar_pc2, axs[1])
+    for ax in axs.flat:
+        ax.set(xlabel='Diffusion measures', ylabel='Loadings')
+    for ax in axs.flat:
+        ax.label_outer()
+    plt.savefig(f'{args.out_folder}/contribution.pdf')
+
+    # Extract the derived newly computed measures from the PCA analysis.
+    logging.info('Saving matrices for PC with eigenvalues > 1...')
+    out = pca.transform(x)
+    out = restore_zero_values(x, out)
+    out = out.swapaxes(0, 1).reshape(len(args.metrics), len(subjects), mat_shape[0], mat_shape[1])
+
+    # Save matrix for significant components
+    nb_pc = eigenvalues[eigenvalues >= 1]
+    for i in range(0, len(nb_pc)):
+        for s in range(0, len(subjects)):
+            save_matrix_in_any_format(f'{args.out_folder}/{subjects[s]}_PC{i+1}.npy',
+                                      out[i, s, :, :])
+
+
+if __name__ == "__main__":
+    main()

scripts/tests/test_compute_pca.py

@@ -0,0 +1,30 @@
+#!/usr/bin/env python3
+# -*- coding: utf-8 -*-
+
+import os
+import tempfile
+
+from scilpy.io.fetcher import get_testing_files_dict, fetch_data, get_home
+
+
+# If they already exist, this only takes 5 seconds (check md5sum)
+fetch_data(get_testing_files_dict(), keys=['stats.zip'])
+tmp_dir = tempfile.TemporaryDirectory()
+
+
+def test_help_option(script_runner):
+    ret = script_runner.run('scil_compute_pca.py',
+                            '--help')
+    assert ret.success
+
+
+def test_execution_pca(script_runner):
+    os.chdir(os.path.expanduser(tmp_dir.name))
+    input_folder = os.path.join(get_home(), 'pca')
+    output_folder = os.path.join(get_home(), 'pca_out')
+    ids = os.path.join(get_home(), 'pca',
+                       'list_id.txt')
+    ret = script_runner.run('scil_compute_pca.py', input_folder, output_folder, '--metrics', 'ad',
+                            'fa', 'md', 'rd', 'nufo', 'afd_total', 'afd_fixel', '--list_ids',
+                            ids, '-f')
+    assert ret.success