Voxel-Based Morphometry on Oasis dataset

This example uses Voxel-Based Morphometry (VBM) to study the relationship between aging and gray matter density.

The data come from the OASIS project. If you use it, you need to agree with the data usage agreement available on the website.

It has been run through a standard VBM pipeline (using SPM8 and NewSegment) to create VBM maps, which we study here.

Predictive modeling analysis: VBM bio-markers of aging?

We run a standard SVM-ANOVA nilearn pipeline to predict age from the VBM data. We use only 100 subjects from the OASIS dataset to limit the memory usage.

Note that for an actual predictive modeling study of aging, the study should be ran on the full set of subjects. Also, all parameters should be set by cross-validation. This includes the smoothing applied to the data and the number of features selected by the ANOVA step. Indeed, even these data-preparation parameter impact significantly the prediction score.

Also, parameters such as the smoothing should be applied to the data and the number of features selected by the ANOVA step should be set by nested cross-validation, as they impact significantly the prediction score.

Brain mapping with mass univariate

SVM weights are very noisy, partly because heavy smoothing is detrimental for the prediction here. A standard analysis using mass-univariate GLM (here permuted to have exact correction for multiple comparisons) gives a much clearer view of the important regions.

---

Note

If you are using Nilearn with a version older than 0.9.0, then you should either upgrade your version or import maskers from the input_data module instead of the maskers module.

That is, you should manually replace in the following example all occurrences of:

from nilearn.maskers import NiftiMasker

with:

from nilearn.input_data import NiftiMasker

import numpy as np

from nilearn import datasets
from nilearn.image import get_data
from nilearn.maskers import NiftiMasker

n_subjects = 100  # more subjects requires more memory

Load Oasis dataset

oasis_dataset = datasets.fetch_oasis_vbm(
    n_subjects=n_subjects, legacy_format=False
)
gray_matter_map_filenames = oasis_dataset.gray_matter_maps
age = oasis_dataset.ext_vars["age"].values

# Split data into training set and test set
from sklearn.model_selection import train_test_split

gm_imgs_train, gm_imgs_test, age_train, age_test = train_test_split(
    gray_matter_map_filenames, age, train_size=0.6, random_state=0
)

# print basic information on the dataset
print(
    "First gray-matter anatomy image (3D) is located at: "
    f"{oasis_dataset.gray_matter_maps[0]}"
)
print(
    "First white-matter anatomy image (3D) is located at: "
    f"{oasis_dataset.white_matter_maps[0]}"
)

First gray-matter anatomy image (3D) is located at: /home/himanshu/nilearn_data/oasis1/OAS1_0001_MR1/mwrc1OAS1_0001_MR1_mpr_anon_fslswapdim_bet.nii.gz
First white-matter anatomy image (3D) is located at: /home/himanshu/nilearn_data/oasis1/OAS1_0001_MR1/mwrc2OAS1_0001_MR1_mpr_anon_fslswapdim_bet.nii.gz

Preprocess data

nifti_masker = NiftiMasker(
    standardize=False, smoothing_fwhm=2, memory="nilearn_cache"
)  # cache options
gm_maps_masked = nifti_masker.fit_transform(gm_imgs_train)

# The features with too low between-subject variance are removed using
# :class:`sklearn.feature_selection.VarianceThreshold`.
from sklearn.feature_selection import VarianceThreshold

variance_threshold = VarianceThreshold(threshold=0.01)
variance_threshold.fit_transform(gm_maps_masked)

# Then we convert the data back to the mask image in order to use it for
# decoding process
mask = nifti_masker.inverse_transform(variance_threshold.get_support())

Prediction pipeline with ANOVA and SVR using nilearn.decoding.DecoderRegressor Object

In nilearn we can benefit from the built-in DecoderRegressor object to do ANOVA with SVR instead of manually defining the whole pipeline. This estimator also uses Cross Validation to select best models and ensemble them. Furthermore, you can pass n_jobs=<some_high_value> to the DecoderRegressor class to take advantage of a multi-core system. To save time (because these are anat images with many voxels), we include only the 1-percent voxels most correlated with the age variable to fit. We also want to set mask hyperparameter to be the mask we just obtained above.

from nilearn.decoding import DecoderRegressor

decoder = DecoderRegressor(
    estimator="svr",
    mask=mask,
    scoring="neg_mean_absolute_error",
    screening_percentile=1,
    n_jobs=2,
    standardize="zscore_sample",
)
# Fit and predict with the decoder
decoder.fit(gm_imgs_train, age_train)

# Sort test data for better visualization (trend, etc.)
perm = np.argsort(age_test)[::-1]
age_test = age_test[perm]
gm_imgs_test = np.array(gm_imgs_test)[perm]
age_pred = decoder.predict(gm_imgs_test)

prediction_score = -np.mean(decoder.cv_scores_["beta"])

print("=== DECODER ===")
print(f"explained variance for the cross-validation: {prediction_score:f}")
print()

=== DECODER ===
explained variance for the cross-validation: 10.670599

Visualization

weight_img = decoder.coef_img_["beta"]

# Create the figure
from nilearn.plotting import plot_stat_map, show

bg_filename = gray_matter_map_filenames[0]
z_slice = 0
display = plot_stat_map(
    weight_img, bg_img=bg_filename, display_mode="z", cut_coords=[z_slice]
)
display.title("SVM weights")
show()

Visualize the quality of predictions

import matplotlib.pyplot as plt

plt.figure(figsize=(6, 4.5))
plt.suptitle(f"Decoder: Mean Absolute Error {prediction_score:.2f} years")
linewidth = 3
plt.plot(age_test, label="True age", linewidth=linewidth)
plt.plot(age_pred, "--", c="g", label="Predicted age", linewidth=linewidth)
plt.ylabel("age")
plt.xlabel("subject")
plt.legend(loc="best")
plt.figure(figsize=(6, 4.5))
plt.plot(
    age_test - age_pred, label="True age - predicted age", linewidth=linewidth
)
plt.xlabel("subject")
plt.legend(loc="best")

• 
• 

<matplotlib.legend.Legend object at 0x7214d93bb980>

Inference with massively univariate model

print("Massively univariate model")

gm_maps_masked = NiftiMasker().fit_transform(gray_matter_map_filenames)
data = variance_threshold.fit_transform(gm_maps_masked)

# Statistical inference
from nilearn.mass_univariate import permuted_ols

# This can be changed to use more CPUs.
neg_log_pvals, t_scores_original_data, _ = permuted_ols(
    age,
    data,  # + intercept as a covariate by default
    n_perm=2000,  # 1,000 in the interest of time; 10000 would be better
    verbose=1,  # display progress bar
    n_jobs=2,
)
signed_neg_log_pvals = neg_log_pvals * np.sign(t_scores_original_data)
signed_neg_log_pvals_unmasked = nifti_masker.inverse_transform(
    variance_threshold.inverse_transform(signed_neg_log_pvals)
)

# Show results
threshold = -np.log10(0.1)  # 10% corrected

fig = plt.figure(figsize=(5.5, 7.5), facecolor="k")

display = plot_stat_map(
    signed_neg_log_pvals_unmasked,
    bg_img=bg_filename,
    threshold=threshold,
    cmap=plt.cm.RdBu_r,
    display_mode="z",
    cut_coords=[z_slice],
    figure=fig,
)
title = (
    "Negative $\\log_{10}$ p-values" "\n(Non-parametric + max-type correction)"
)
display.title(title)

n_detections = (get_data(signed_neg_log_pvals_unmasked) > threshold).sum()
print(f"\n{int(n_detections)} detections")

show()

Massively univariate model
[Parallel(n_jobs=2)]: Using backend LokyBackend with 2 concurrent workers.
[Parallel(n_jobs=2)]: Done   2 out of   2 | elapsed:   33.1s finished

1966 detections

Estimated memory usage: 2486 MB