def merge_clusters(labels_a_weights, n, labels_b, k):
"""
The merge strategy is described in Section C: Voting Procedure in the paper
Detecting Regions of Interest in fMRI, found in the link
http://ieeexplore.ieee.org/stamp/stamp.jsp?arnumber=1006726
Parameters
----------
labels_a_weights : 3D np array: [labels_3d_index_x, labels_3d_index_y, labels_3d_index_z]. This
is the weights of the primary volume. This is the outcome of merging n volumes of labels together
n : no. of volumes already merged into labels_a
labels_b : 3D np array: [labels_3d_index_x, labels_3d_index_y, labels_3d_index_z]. We merge
this volume of labels into labels_a
k : no. of clusters
Returns
-------
labels_3d : array has the same shape as img_data.shape. Each elem is the
cluster label of the merged data.
"""
assert labels_a_weights.shape[1:] == labels_b.shape
shape = labels_b.shape
# separate all points to their respective clusters
cluster_to_point_map_b = [set() for i in range(k)]
for i in vol_index_iter(shape):
cluster_to_point_map_b[labels_b[i]].add(i)
# compute a k * k matrix A with A[i, j] indicating similarity between cluster i in a and cluster j in b
inter_cluster_similarity = np.zeros((k, k)) - 1
for a_i, b_i in product(range(k), range(k)):
if a_i != b_i:
inter_cluster_similarity[(a_i, b_i)] = weighted_similarity(
labels_a_weights[a_i], cluster_to_point_map_b[b_i])
# based on the inter-cluster-similarity matrix, decide cluster mapping, matching the most similar a-b pair of clusters first.
cluster_map = {}
while len(cluster_map) < k:
a_index, b_index = np.unravel_index(
np.argmax(inter_cluster_similarity), (k, k))
cluster_map[a_index] = b_index
inter_cluster_similarity[a_index, :] = -float('inf')
inter_cluster_similarity[:, b_index] = -float('inf')
# update weights on the primary volume
for i, j in cluster_map.items():
for index in vol_index_iter(shape):
if labels_a_weights[i][index] == 0: continue
labels_a_weights[i][index] += n / (n + 1)
new_cluster_index = i if index in cluster_to_point_map_b[
j] else find_cluster(cluster_to_point_map_b, index)
labels_a_weights[new_cluster_index][index] += 1 / (n + 1)
def merge_clusters(labels_a_weights, n, labels_b, k):
"""
The merge strategy is described in Section C: Voting Procedure in the paper
Detecting Regions of Interest in fMRI, found in the link
http://ieeexplore.ieee.org/stamp/stamp.jsp?arnumber=1006726
Parameters
----------
labels_a_weights : 3D np array: [labels_3d_index_x, labels_3d_index_y, labels_3d_index_z]. This
is the weights of the primary volume. This is the outcome of merging n volumes of labels together
n : no. of volumes already merged into labels_a
labels_b : 3D np array: [labels_3d_index_x, labels_3d_index_y, labels_3d_index_z]. We merge
this volume of labels into labels_a
k : no. of clusters
Returns
-------
labels_3d : array has the same shape as img_data.shape. Each elem is the
cluster label of the merged data.
"""
assert labels_a_weights.shape[1:] == labels_b.shape
shape = labels_b.shape
# separate all points to their respective clusters
cluster_to_point_map_b = [set() for i in range(k)]
for i in vol_index_iter(shape):
cluster_to_point_map_b[labels_b[i]].add(i)
# compute a k * k matrix A with A[i, j] indicating similarity between cluster i in a and cluster j in b
inter_cluster_similarity = np.zeros((k, k)) - 1
for a_i, b_i in product(range(k), range(k)):
if a_i != b_i:
inter_cluster_similarity[(a_i, b_i)] = weighted_similarity(labels_a_weights[a_i], cluster_to_point_map_b[b_i])
# based on the inter-cluster-similarity matrix, decide cluster mapping, matching the most similar a-b pair of clusters first.
cluster_map = {}
while len(cluster_map) < k:
a_index, b_index = np.unravel_index(np.argmax(inter_cluster_similarity), (k, k))
cluster_map[a_index] = b_index
inter_cluster_similarity[a_index,:] = -float('inf')
inter_cluster_similarity[:, b_index] = -float('inf')
# update weights on the primary volume
for i, j in cluster_map.items():
for index in vol_index_iter(shape):
if labels_a_weights[i][index] == 0: continue
labels_a_weights[i][index] += n / (n + 1)
new_cluster_index = i if index in cluster_to_point_map_b[j] else find_cluster(cluster_to_point_map_b, index)
labels_a_weights[new_cluster_index][index] += 1 / (n + 1)
def pad_boundary_per_image(data, brain_mask, pad_thickness):
brain_points, non_brain_points = [], []
for i in vol_index_iter(data.shape):
if brain_mask[i]:
brain_points.append(i)
else:
non_brain_points.append(i)
tree = cKDTree(np.array(brain_points))
non_brain_points = [i for i in vol_index_iter(data.shape) if (not brain_mask[i])]
neighbors_list = tree.query_ball_point(non_brain_points, pad_thickness)
for list_i, neighbors in enumerate(neighbors_list):
i = non_brain_points[list_i]
if len(neighbors) != 0:
data[i] = np.mean([data[brain_points[j]] for j in neighbors])
def pad_boundary_per_image(data, brain_mask, pad_thickness):
brain_points, non_brain_points = [], []
for i in vol_index_iter(data.shape):
if brain_mask[i]:
brain_points.append(i)
else:
non_brain_points.append(i)
tree = cKDTree(np.array(brain_points))
non_brain_points = [i for i in vol_index_iter(data.shape) if (not brain_mask[i])]
neighbors_list = tree.query_ball_point(non_brain_points, pad_thickness)
for list_i, neighbors in enumerate(neighbors_list):
i = non_brain_points[list_i]
if len(neighbors) != 0:
data[i] = np.mean([data[brain_points[j]] for j in neighbors])
def __init__(self, in_brain_mask, dist_from_center):
points = [
i for i in vol_index_iter(in_brain_mask.shape) if in_brain_mask[i]
]
self.tree = cKDTree(np.array(points))
self.points = points
self.dist_from_center = dist_from_center
def weighted_similarity(cluster_a_weights, cluster_b):
shape = cluster_a_weights.shape
accu_weight = 0
for i in vol_index_iter(shape):
if i in cluster_b:
accu_weight += cluster_a_weights[i]
return accu_weight / np.sum(cluster_a_weights)
def weighted_similarity(cluster_a_weights, cluster_b):
shape = cluster_a_weights.shape
accu_weight = 0
for i in vol_index_iter(shape):
if i in cluster_b:
accu_weight += cluster_a_weights[i]
return accu_weight / np.sum(cluster_a_weights)
def correlation_map(data, cond_filename):
tr_times = np.arange(0, 30, TR)
convolved = conv_main(data.shape[-1], cond_filename, TR)
corrs = np.zeros((data.shape[:-1]))
for i in general_utils.vol_index_iter(data.shape[:-1]):
corrs[i] = np.corrcoef(data[i], convolved)[1,0]
return corrs
def correlation_map(data, cond_filename):
tr_times = np.arange(0, 30, TR)
convolved = conv_main(data.shape[-1], cond_filename, TR)
corrs = np.zeros((data.shape[:-1]))
for i in general_utils.vol_index_iter(data.shape[:-1]):
corrs[i] = np.corrcoef(data[i], convolved)[1, 0]
return corrs
def correlation_map_without_convoluation(data, cond_filename):
n_trs = data.shape[-1] + 5
time_course = events2neural_rounded(cond_filename, TR, n_trs)
time_course=time_course[5:]
correlations = np.zeros(data.shape[:-1])
for i in general_utils.vol_index_iter(data.shape[:-1]):
vox_values = data[i]
correlations[i] = np.corrcoef(time_course, vox_values)[1, 0]
def correlation_map_without_convoluation(data, cond_filename):
n_trs = data.shape[-1] + 5
time_course = events2neural_rounded(cond_filename, TR, n_trs)
time_course = time_course[5:]
correlations = np.zeros(data.shape[:-1])
for i in general_utils.vol_index_iter(data.shape[:-1]):
vox_values = data[i]
correlations[i] = np.corrcoef(time_course, vox_values)[1, 0]
def plot_cluster_3d(labels_3d, cluster_index):
Xs, Ys, Zs = [], [], []
for i, j, k in vol_index_iter(labels_3d.shape):
if labels_3d[i, j, k] == cluster_index:
Xs.append(i)
Ys.append(j)
Zs.append(k)
fig = plt.figure()
ax = Axes3D(fig)
ax.scatter(Xs, Ys, Zs)
ax.show()
def plot_cluster_3d(labels_3d, cluster_index):
Xs, Ys, Zs = [], [], []
for i, j, k in vol_index_iter(labels_3d.shape):
if labels_3d[i, j, k] == cluster_index:
Xs.append(i)
Ys.append(j)
Zs.append(k)
fig = plt.figure()
ax = Axes3D(fig)
ax.scatter(Xs, Ys, Zs)
ax.show()
def correlation_map(data, cond_filename):
"""
Generate a cross-correlation per voxel between BOLD signals and a convolved
gamma baseline function. Assume that the first 5 images are already dropped.
"""
convolved = conv_main(data.shape[-1] + 5, cond_filename, TR)[5:]
corrs = np.zeros((data.shape[:-1]))
for i in gu.vol_index_iter(data.shape[:-1]):
r = np.corrcoef(data[i], convolved)[1,0]
if np.isnan(r):
r = 0
corrs[i] = r
return corrs
def correlation_map(data, cond_filename):
"""
Generate a cross-correlation per voxel between BOLD signals and a convolved
gamma baseline function. Assume that the first 5 images are already dropped.
"""
convolved = conv_main(data.shape[-1] + 5, cond_filename, TR)[5:]
corrs = np.zeros((data.shape[:-1]))
for i in gu.vol_index_iter(data.shape[:-1]):
r = np.corrcoef(data[i], convolved)[1, 0]
if np.isnan(r):
r = 0
corrs[i] = r
return corrs
def correlation_map_without_convoluation(data, cond_filename):
"""
Generate a cross-correlation per voxel between BOLD signals and a square wave
baseline function, which represents the boolean array of the on-off time
course. Assume that the first 5 images are already dropped.
"""
n_trs = data.shape[-1] + 5
time_course = events2neural_rounded(cond_filename, TR, n_trs)[5:]
correlations = np.zeros(data.shape[:-1])
for i in gu.vol_index_iter(data.shape[:-1]):
vox_values = data[i]
r = np.corrcoef(time_course, vox_values)[1, 0]
if np.isnan(r):
r = 0
correlations[i] = r
return correlations
def correlation_map_without_convoluation(data, cond_filename):
"""
Generate a cross-correlation per voxel between BOLD signals and a square wave
baseline function, which represents the boolean array of the on-off time
course. Assume that the first 5 images are already dropped.
"""
n_trs = data.shape[-1] + 5
time_course = events2neural_rounded(cond_filename, TR, n_trs)[5:]
correlations = np.zeros(data.shape[:-1])
for i in gu.vol_index_iter(data.shape[:-1]):
vox_values = data[i]
r = np.corrcoef(time_course, vox_values)[1, 0]
if np.isnan(r):
r = 0
correlations[i] = r
return correlations
def __init__(self, in_brain_mask, dist_from_center): points = [i for i in vol_index_iter(in_brain_mask.shape) if in_brain_mask[i]] self.tree = cKDTree(np.array(points)) self.points = points self.dist_from_center = dist_from_center
def assign_points(labels_weights, shape):
labels = np.zeros((shape))
for x, y, z in vol_index_iter(shape):
labels[x, y, z] = np.argmax(labels_weights[:, x, y, z])
return labels
def form_initial_weights(labels, n_clusters):
shape = labels.shape
result_labels_weights = np.zeros((n_clusters, ) + shape)
for i, j, k in vol_index_iter(shape):
result_labels_weights[labels[i, j, k], i, j, k] = 1
return result_labels_weights
def form_initial_weights(labels, n_clusters):
shape = labels.shape
result_labels_weights = np.zeros((n_clusters,) + shape)
for i, j, k in vol_index_iter(shape):
result_labels_weights[labels[i, j, k], i, j, k] = 1
return result_labels_weights
def assign_points(labels_weights, shape):
labels = np.zeros((shape))
for x,y,z in vol_index_iter(shape):
labels[x,y,z] = np.argmax(labels_weights[:,x,y,z])
return labels
