def extract_super_pixels(mat_adj=None,
test_left=None,
test_right=None,
mat_cat=None,
num_neighbors=8,
cor_choice='mean',
connectivity=None,
min_pixels=50,
image=None,
plot=False,
use_mean_image=False):
if image is None:
if mat_cat is None:
mat_cat = torch.cat([m[test_left:test_right] for m in mat_adj],
dim=0)
cor_global = neighbor_cor(mat_cat,
neighbors=num_neighbors,
plot=plot,
choice=cor_choice,
title='correlation map')
if use_mean_image:
cor_global = mat_cat.mean(0) * cor_global
cor_global = cor_global / cor_global.max()
image = cor_global.detach().cpu().numpy()
else:
cor_global = image # for backward compatibility
if isinstance(image, torch.Tensor):
image = image.detach().cpu().numpy()
label_image, regions = get_label_image(image,
min_pixels=min_pixels,
connectivity=connectivity,
plot=False)
if plot:
plot_image_label_overlay(image, label_image)
return cor_global, label_image, regions
def detrend_high_magnification(mat, skip_segments=1, num_segments=6, period=500, train_size_left=0, train_size_right=350,
linear_order=3, plot=False, signal_start=0, signal_end=100, filepath=None, size=(-1, 180, 300),
device=torch.device('cuda'), start0=None, end0=None, return_mat=False, **kwargs):
if mat is None:
mat = load_file(filepath=filepath, size=size, dtype=np.uint16, device=device)
L, nrow, ncol = mat.size()
if period == 'unknown':
period = L
if signal_end == 'period':
signal_end = period
train_idx = ([range(skip_segments*period)] +
[range(i*period, train_size_left+i*period) for i in range(skip_segments, num_segments)] +
[range((i+1)*period - train_size_right, (i+1)*period) for i in range(skip_segments, num_segments)])
train_idx = functools.reduce(lambda x,y: list(x)+list(y), train_idx)
input_aug = torch.linspace(-2, 2, L, device=device)
beta, trend = linear_regression(X=input_aug[train_idx], Y=mat.reshape(L, -1)[train_idx], order=linear_order, X_test=input_aug)
mat_adj = mat - trend.reshape(L, nrow, ncol)
if plot:
frame_mean = mat.mean(-1).mean(-1)
plt.figure(figsize=(20, 10))
plt.plot(frame_mean.cpu(), 'o--', linewidth=1, markersize=2)
for i in range(num_segments):
plt.axvline(i*period+signal_start, color='g', linestyle='-.')
plt.axvline(i*period+signal_end, color='g', linestyle='-.')
plt.title('Frame mean intensity')
plt.show()
imshow(mat.mean(0), title='Mean intensity')
imshow(trend.mean(0).reshape(nrow, ncol), title='Trend')
imshow(mat_adj.mean(0).reshape(nrow, ncol), title='Detrended')
cor = neighbor_cor(mat, neighbors=8, plot=True, choice='max', title='cor mat')
plot_hist(cor, show=True)
cor = neighbor_cor(trend.reshape(-1, nrow, ncol), neighbors=8, plot=True, choice='max', title='cor trend')
plot_hist(cor, show=True)
cor = neighbor_cor(mat_adj.reshape(-1, nrow, ncol), neighbors=8, plot=True, choice='max', title='cor mat_adj')
plot_hist(cor, show=True)
plot_mean_intensity(mat, detrended=mat_adj, plot_detrended=True, plot_segments=True, num_frames=L, period=period,
signal_length=signal_end-signal_start)
if start0 is not None and end0 is not None:
mat_adj = [mat_adj[s:e] for s, e in zip(start0, end0)]
if return_mat:
return mat, mat_adj
else:
return mat_adj
def get_trace(mat=None,
mat_adj=None,
seg_idx=None,
submat=None,
weight_percentile=weight_percentile,
plot_singular_values=plot_singular_values):
if submat is None:
if mat is None:
mat = mat_adj[seg_idx]
submat = mat[:, minr:maxr, minc:maxc]
if cor is None:
cor_ = neighbor_cor(submat,
neighbors=8,
plot=False,
choice='max',
title='cor',
return_adj_list=False)
else:
cor_ = cor[minr:maxr, minc:maxc]
if weight_percentile is None:
weight = cor_
else:
weight = nn.functional.threshold(cor_,
np.percentile(
cor_[label_mask.bool()].cpu(),
weight_percentile),
0,
inplace=False)
if weight.sum() == 0:
weight = 1
denominator = (
label_mask *
weight).sum() if weighted_denominator else label_mask.sum()
trace = ((submat * label_mask * weight).sum(-1).sum(-1) / denominator)
if plot_singular_values:
u, s, v = torch.svd(submat.reshape(len(submat), -1))
plot_tensor(s,
marker='o--',
linewidth=1,
figsize=(20, 10),
show=True)
return submat, trace, weight