Analysis of an fMRI dataset with a Finite Impule Response (FIR) model

FIR models are used to estimate the hemodyamic response non-parametrically. The example below shows that they’re good to do statistical inference even on fast event-related fMRI datasets.

Here, we demonstrate the use of a FIR model with 3 lags, computing 4 contrasts from a single subject dataset from the “Neurospin Localizer”. It is a fast event related design: During 5 minutes, 80 events of the following types are presented : [‘audio_computation’, ‘audio_left_hand_button_press’, ‘audio_right_hand_button_press’, ‘horizontal_checkerboard’, ‘sentence_listening’, ‘sentence_reading’, ‘vertical_checkerboard’, ‘visual_computation’, ‘visual_left_hand_button_press’, ‘visual_right_hand_button_press’]

At first, we grab the localizer data.

import pandas as pd

from nilearn.datasets import func
from nilearn.plotting import plot_stat_map

data = func.fetch_localizer_first_level()
fmri_img = data.epi_img
t_r = 2.4
events_file = data["events"]
events = pd.read_table(events_file)

[get_dataset_dir] Dataset created in
/home/runner/nilearn_data/localizer_first_level
[fetch_single_file] Downloading data from https://osf.io/2bqxn/download ...
[_chunk_report_] Downloaded 6291456 of 34532740 bytes (18.2%%,    4.7s
remaining)
[_chunk_report_] Downloaded 13271040 of 34532740 bytes (38.4%%,    3.3s
remaining)
[_chunk_report_] Downloaded 17530880 of 34532740 bytes (50.8%%,    3.0s
remaining)
[_chunk_report_] Downloaded 22503424 of 34532740 bytes (65.2%%,    2.2s
remaining)
[_chunk_report_] Downloaded 29040640 of 34532740 bytes (84.1%%,    1.0s
remaining)
[_chunk_report_] Downloaded 33480704 of 34532740 bytes (97.0%%,    0.2s
remaining)
[fetch_single_file]  ...done. (8 seconds, 0 min)

[uncompress_file] Extracting data from
/home/runner/nilearn_data/localizer_first_level/cac0ef0f4307834e60fedd2a0448862a
/download...
[uncompress_file] .. done.

Next solution is to try Finite Impulse Response (FIR) models: we just say that the hrf is an arbitrary function that lags behind the stimulus onset. In the present case, given that the numbers of conditions is high, we should use a simple FIR model.

Concretely, we set hrf_model to ‘fir’ and fir_delays to [1, 2, 3] (scans) corresponding to a 3-step functions on the [1 * t_r, 4 * t_r] seconds interval.

from nilearn.glm.first_level import FirstLevelModel
from nilearn.plotting import plot_contrast_matrix, plot_design_matrix

first_level_model = FirstLevelModel(t_r, hrf_model="fir", fir_delays=[1, 2, 3])
first_level_model = first_level_model.fit(fmri_img, events=events)
design_matrix = first_level_model.design_matrices_[0]
plot_design_matrix(design_matrix)

<AxesSubplot:label='conditions', ylabel='scan number'>

We have to adapt contrast specification. We characterize the BOLD response by the sum across the three time lags. It’s a bit hairy, sorry, but this is the price to pay for flexibility…

import numpy as np

contrast_matrix = np.eye(design_matrix.shape[1])
contrasts = {
    column: contrast_matrix[i]
    for i, column in enumerate(design_matrix.columns)
}
conditions = events.trial_type.unique()
for condition in conditions:
    contrasts[condition] = np.sum(
        [
            contrasts[name]
            for name in design_matrix.columns
            if name[: len(condition)] == condition
        ],
        0,
    )

contrasts["audio"] = np.sum(
    [
        contrasts[name]
        for name in [
            "audio_right_hand_button_press",
            "audio_left_hand_button_press",
            "audio_computation",
            "sentence_listening",
        ]
    ],
    0,
)
contrasts["video"] = np.sum(
    [
        contrasts[name]
        for name in [
            "visual_right_hand_button_press",
            "visual_left_hand_button_press",
            "visual_computation",
            "sentence_reading",
        ]
    ],
    0,
)

contrasts["computation"] = (
    contrasts["audio_computation"] + contrasts["visual_computation"]
)
contrasts["sentences"] = (
    contrasts["sentence_listening"] + contrasts["sentence_reading"]
)

contrasts = {
    "left-right": (
        contrasts["visual_left_hand_button_press"]
        + contrasts["audio_left_hand_button_press"]
        - contrasts["visual_right_hand_button_press"]
        - contrasts["audio_right_hand_button_press"]
    ),
    "H-V": (
        contrasts["horizontal_checkerboard"]
        - contrasts["vertical_checkerboard"]
    ),
    "audio-video": contrasts["audio"] - contrasts["video"],
    "sentences-computation": (
        contrasts["sentences"] - contrasts["computation"]
    ),
}

Take a look at the contrasts.

plot_contrast_matrix(contrasts["left-right"], design_matrix)

<AxesSubplot:label='conditions'>

Take a breath.

We can now proceed by estimating the contrasts and displaying them.

import matplotlib.pyplot as plt

fig = plt.figure(figsize=(11, 3))
for index, (contrast_id, contrast_val) in enumerate(contrasts.items()):
    ax = plt.subplot(1, len(contrasts), 1 + index)
    z_map = first_level_model.compute_contrast(
        contrast_val, output_type="z_score"
    )
    plot_stat_map(
        z_map,
        display_mode="z",
        threshold=3.0,
        title=contrast_id,
        axes=ax,
        cut_coords=1,
    )
plt.show()

The result is acceptable. Note that we’re asking a lot of questions to a small dataset, yet with a relatively large number of experimental conditions.

Estimated memory usage: 416 MB