def _calculate_cluster_based_test_statistic(dataset_1,
dataset_2,
min_area=1,
cluster_alpha=0.05,
sample_statistic='independent',
cluster_statistic='maxsum'):
if sample_statistic == 'independent':
t_test_result = sstat.ttest_ind(dataset_1,
dataset_2,
axis=-1,
equal_var=False)
if sample_statistic == 'dependent':
t_test_result = sstat.ttest_rel(dataset_1, dataset_2, axis=-1)
masked_pvalues = np.zeros(t_test_result.pvalue.shape)
masked_pvalues[t_test_result.pvalue < cluster_alpha] = 1
cluster_labels, num = measurements.label(masked_pvalues)
cluster_areas, _ = measurements._stats(
masked_pvalues,
cluster_labels,
index=np.arange(cluster_labels.max() + 1))
clusters_over_min_area = np.squeeze(np.argwhere(cluster_areas > min_area))
cluster_labels_over_min_area = np.ones(cluster_labels.shape) * -1
cluster_statistic_result = []
if clusters_over_min_area.shape == ():
clusters_over_min_area = [clusters_over_min_area]
for c_index in np.arange(len(clusters_over_min_area)):
cluster = clusters_over_min_area[c_index]
if cluster_statistic == 'maxsum':
cluster_statistic_result.append(
np.sum(t_test_result.pvalue[cluster_labels == cluster]))
if cluster_statistic == 'maxarea':
cluster_statistic_result.append(cluster_areas[cluster])
cluster_labels_over_min_area[cluster_labels == cluster] = c_index
return cluster_statistic_result, cluster_labels_over_min_area
def vertices_to_parcels(data, *, lhannot, rhannot, drop=None):
"""
Reduces vertex-level `data` to parcels defined in annotation files
Takes average of vertices within each parcel, excluding np.nan values
(i.e., np.nanmean). Assigns np.nan to parcels for which all vertices are
np.nan.
Parameters
----------
data : (N,) numpy.ndarray
Vertex-level data to be reduced to parcels
{lh,rh}annot : str
Path to .annot file containing labels to parcels on the {left,right}
hemisphere
drop : list, optional
Specifies regions in {lh,rh}annot that should be removed from the
parcellated version of `data`. If not specified, vertices corresponding
to parcels defined in `netneurotools.freesurfer.FSIGNORE` will be
removed. Default: None
Reurns
------
reduced : numpy.ndarray
Parcellated `data`, without regions specified in `drop`
"""
if drop is None:
drop = FSIGNORE
drop = _decode_list(drop)
data = np.vstack(data)
n_parc = expected = 0
for a in [lhannot, rhannot]:
vn, _, names = read_annot(a)
expected += len(vn)
names = _decode_list(names)
n_parc += len(names) - len(set(drop) & set(names))
if expected != len(data):
raise ValueError('Number of vertices in provided annotation files '
'differs from size of vertex-level data array.\n'
' EXPECTED: {} vertices\n'
' RECEIVED: {} vertices'.format(
expected, len(data)))
reduced = np.zeros((n_parc, data.shape[-1]), dtype=data.dtype)
start = end = n_parc = 0
for annot in [lhannot, rhannot]:
# read files and update end index for `data`
labels, ctab, names = read_annot(annot)
names = _decode_list(names)
indices = np.unique(labels)
end += len(labels)
for idx in range(data.shape[-1]):
# get average of vertex-level data within parcels
# set all NaN values to 0 before calling `_stats` because we are
# returning sums, so the 0 values won't impact the sums (if we left
# the NaNs then all parcels with even one NaN entry would be NaN)
currdata = np.squeeze(data[start:end, idx])
isna = np.isnan(currdata)
counts, sums = _stats(np.nan_to_num(currdata), labels, indices)
# however, we do need to account for the NaN values in the counts
# so that our means are similar to what we'd get from e.g.,
# np.nanmean here, our "sums" are the counts of NaN values in our
# parcels
_, nacounts = _stats(isna, labels, indices)
counts = (np.asanyarray(counts, dtype=float) -
np.asanyarray(nacounts, dtype=float))
with np.errstate(divide='ignore', invalid='ignore'):
currdata = sums / counts
# get indices of unkown and corpuscallosum and delete from parcels
inds = sorted([names.index(f) for f in set(drop) & set(names)])
currdata = np.delete(currdata, inds)
# store parcellated data
reduced[n_parc:n_parc + len(names) - len(inds), idx] = currdata
start = end
n_parc += len(names) - len(inds)
return np.squeeze(reduced)
def vertices_to_parcels(data, *, lhannot, rhannot, drop=None):
"""
Reduces vertex-level `data` to parcels defined in annotation files
Takes average of vertices within each parcel, excluding np.nan values
(i.e., np.nanmean). Assigns np.nan to parcels for which all vertices are
np.nan.
Parameters
----------
data : (N,) numpy.ndarray
Vertex-level data to be reduced to parcels
{lh,rh}annot : str
Path to .annot file containing labels to parcels on the {left,right}
hemisphere
drop : list, optional
Specifies regions in {lh,rh}annot that should be removed from the
parcellated version of `data`. If not specified, 'unknown' and
'corpuscallosum' will be removed. Default: None
Reurns
------
reduced : numpy.ndarray
Parcellated `data`, without regions specified in `drop`
"""
if drop is None:
drop = ['unknown', 'corpuscallosum']
drop = _decode_list(drop)
start = end = 0
reduced = []
# check this so we're not unduly surprised by anything...
expected = sum([len(read_annot(a)[0]) for a in [lhannot, rhannot]])
if expected != len(data):
raise ValueError('Number of vertices in provided annotation files '
'differs from size of vertex-level data array.\n'
' EXPECTED: {} vertices\n'
' RECEIVED: {} vertices'.format(
expected, len(data)))
for annot in [lhannot, rhannot]:
# read files and update end index for `data`
labels, ctab, names = read_annot(annot)
names = _decode_list(names)
indices = np.unique(labels)
end += len(labels)
# get average of vertex-level data within parcels
# set all NaN values to 0 before calling `_stats` because we are
# returning sums, so the 0 values won't impact the sums (if we left
# the NaNs then all parcels with even one NaN entry would be NaN)
currdata = np.squeeze(data[start:end])
isna = np.isnan(currdata)
counts, sums = _stats(np.nan_to_num(currdata), labels, indices)
# however, we do need to account for the NaN values in the counts
# so that our means are similar to what we'd get from e.g., np.nanmean
# here, our "sums" are the counts of NaN values in our parcels
_, nacounts = _stats(isna, labels, indices)
counts = (np.asanyarray(counts, dtype=float) -
np.asanyarray(nacounts, dtype=float))
with np.errstate(divide='ignore', invalid='ignore'):
currdata = sums / counts
# get indices of unkown and corpuscallosum and delete from parcels
inds = [names.index(f) for f in drop]
currdata = np.delete(currdata, inds)
# store parcellated data
reduced.append(currdata)
start = end
return np.hstack(reduced)
def reduce_from_vertices(data, rhannot, lhannot):
"""
Reduces vertex-level `data` to parcels defined in annotation files
Takes average of vertices within each parcel, excluding np.nan values
(i.e., np.nanmean). Assigns np.nan to parcels for which all vertices are
np.nan.
Parameters
----------
data : (N,) numpy.ndarray
Vertex-level data to be reduced to parcels
{rh,lh}annot : str
Path to .annot file containing labels to parcels on the {right,left}
hemisphere
Reurns
------
reduced : numpy.ndarray
Parcellated data
"""
drop = [b'unknown', b'corpuscallosum']
start = end = 0
reduced = []
# check this so we're not unduly surprised by anything...
expected = sum([len(read_annot(a)[0]) for a in [rhannot, lhannot]])
if expected != len(data):
raise ValueError('Number of vertices in provided annotation files '
'differs from size of vertex-level data array.\n'
' EXPECTED: {} vertices\n'
' RECEIVED: {} vertices'.format(
expected, len(data)))
for annot in [rhannot, lhannot]:
# read files and update end index for `data`
labels, ctab, names = read_annot(annot)
end += len(labels)
# get average of vertex-level data within parcels
# set all NaN values to 0 before calling `_stats` because we are
# returning sums, so the 0 values won't impact the sums (if we left
# the NaNs then all parcels with a single NaN value would be NaN)
currdata = data[start:end].copy()
isna = np.isnan(currdata)
currdata[isna] = 0
counts, sums = _stats(currdata, labels, np.unique(labels))
# however, we do need to account for the NaN values in the counts
# so that our means are similar to what we'd get from e.g., np.nanmean
# here, our "sums" are the counts of NaN values in our parcels
_, nacounts = _stats(isna, labels, np.unique(labels))
counts = (np.asanyarray(counts).astype(float) -
np.asanyarray(nacounts).astype(float))
with np.errstate(divide='ignore', invalid='ignore'):
currdata = sums / counts
# get indices of unkown and corpuscallosum and delete from parcels
inds = [names.index(f) for f in drop]
currdata = np.delete(currdata, inds)
# store parcellated data
reduced.append(currdata)
start = end
return np.hstack(reduced)
def image_segmentation(ndvi, predict):
write_cog(ndvi.to_array().compute(), "NDVI.tif", overwrite=True)
# store temp files somewhere
directory = "tmp"
if not os.path.exists(directory):
os.mkdir(directory)
tmp = "tmp/"
# inputs to image seg
tiff_to_segment = "NDVI.tif"
kea_file = "NDVI.kea"
segmented_kea_file = "segmented.kea"
# convert tiff to kea
gdal.Translate(
destName=kea_file, srcDS=tiff_to_segment, format="KEA", outputSRS="EPSG:6933"
)
# run image seg
with HiddenPrints():
segutils.runShepherdSegmentation(
inputImg=kea_file,
outputClumps=segmented_kea_file,
tmpath=tmp,
numClusters=60,
minPxls=100,
)
# convert kea to tif
kwargs = {
'outputType': gdal.GDT_Float32,
}
gdal.Translate(
destName=segmented_kea_file[:-3]+'tif',
srcDS=segmented_kea_file,
outputSRS="EPSG:6933",
format='GTiff',
**kwargs
)
# open segments
segments = xr.open_rasterio(segmented_kea_file[:-3]+'tif').squeeze().values
# calculate mode
count, _sum = _stats(predict, labels=segments, index=segments)
mode = _sum > (count / 2)
mode = xr.DataArray(
mode, coords=predict.coords, dims=predict.dims, attrs=predict.attrs
)
# remove the tmp folder
shutil.rmtree(tmp)
os.remove(kea_file)
os.remove(segmented_kea_file)
os.remove(tiff_to_segment)
os.remove(segmented_kea_file[:-3]+'tif')
return mode.chunk({})
def sum(input, labels=None, index=None):
count, sum = _stats(input, labels, index)
return sum
