def test_gaussian_coord_smoothing():
coords = [(0.0, 0.0, 0.0), (10.0, -10.0, 30.0)]
computed_img = img_utils.gaussian_coord_smoothing(coords)
masker = NiftiSpheresMasker(coords + [(-10.0, 10.0, -30)]).fit()
values = masker.transform(computed_img)[0]
assert (values[:2] > computed_img.get_data().max() / 2.0).all()
assert values[-1] == pytest.approx(0.0)
def test_coordinates_to_maps():
coords = pd.DataFrame.from_dict({
"pmid": [3, 17, 17, 2, 2],
"x": [0.0, 0.0, 10.0, 5.0, 3.0],
"y": [0.0, 0.0, -10.0, 15.0, -9.0],
"z": [27.0, 0.0, 30.0, 17.0, 177.0],
})
maps, masker = img_utils.coordinates_to_maps(coords)
assert maps.shape == (3, 28542)
coords_17 = [(0.0, 0.0, 0.0), (10.0, -10.0, 30.0)]
img_17 = img_utils.gaussian_coord_smoothing(coords_17, target_affine=4.0)
assert np.allclose(masker.transform(img_17),
maps.loc[17, :].values,
atol=1e-10)
help=".csv file containing the coordinates. Must have columns"
" 'pmid', 'x', 'y', and 'z'.",
)
parser.add_argument("output_directory",
help="directory where generated maps are saved.")
parser.add_argument("--fwhm",
type=float,
default=8.0,
help="full width at half maximum")
args = parser.parse_args()
out_dir = pathlib.Path(args.output_directory)
out_dir.mkdir(parents=True, exist_ok=True)
coordinates = pd.read_csv(args.coordinates_csv)
articles = coordinates.groupby("pmid")
for i, (pmid, article_coordinates) in enumerate(articles):
print(
"{:.1%} pmid: {:< 20}".format(i / len(articles), pmid),
end="\r",
flush=True,
)
img_file = out_dir / "pmid_{}.nii.gz".format(pmid)
if not img_file.is_file():
img = img_utils.gaussian_coord_smoothing(
article_coordinates.loc[:, ["x", "y", "z"]].values, fwhm=args.fwhm)
img.to_filename(str(img_file))
print("\n")
print(out_dir)
# ------------------------------
# We load example subject-level tmaps from a localizer dataset, and also
# generate a brain maps from a set of MNI coordinates.
queries = {}
contrasts = ["left vs right button press", "sentence listening"]
for contrast in contrasts:
query_map = fetch_localizer_contrasts([contrast],
n_subjects=1,
get_tmaps=True)["tmaps"][0]
queries[contrast] = query_map
dmn_coords = [(0, -52, 18), (-46, -68, 32), (46, -68, 32), (1, 50, -5)]
dmn_img = gaussian_coord_smoothing(dmn_coords, encoder.get_masker())
masked_dmn = encoder.get_masker().transform(dmn_img).ravel()
queries["DMN coordinates"] = dmn_img
######################################################################
# Discover which terms have activations similar to the query map
# --------------------------------------------------------------
# Here we simply use the dot product with the absolute values of the input map
for name, query_map in queries.items():
masked_query = encoder.get_masker().transform(query_map).ravel()
similarities = np.abs(masked_query).dot(term_maps.T)
# rescale by document frequencies, optional
similarities *= np.log(1 + encoder.document_frequencies().values.ravel())
top_20 = np.argsort(similarities)[::-1][:20]
top_terms = voc[top_20].ravel()