def measure_max(source,
points,
search,
radii,
sink=None,
processes=None,
verbose=False):
"""Find index in local search indices with a voxel with value smaller than a specified value for a list of points.
Arguments
---------
source : array
Data source.
points : array
List of linear indices of center points.
search : array
List of linear indices to add to the center index defining the local search area.
radii : array
The maximal index in the search array for each point to use.
sink : array or None
Optional sink for result indices.
processes : int or None
Number of processes to use.
verbose : bool
If True, print progress info.
Returns
-------
sink : array
Linear array with length of points containing the first search index with voxel below value.
"""
processes, timer = ap._initialize_processing(processes=processes,
verbose=verbose,
function='measure_max')
source1d = ap._initialize_source(source, as_1d=True)
sink = ap._initialize_sink(sink=sink,
shape=points.shape,
dtype=source.dtype)
if sink.shape != points.shape:
raise RuntimeError('Sink has invalid size %r not %r' %
(points.shape, points.shape))
#print(source1d.dtype, points.dtype, search.dtype, type(value), sink.dtype);
#print(source1d.shape, points.shape, sink.shape)
print source1d, points, search, radii, sink
code.measure_max(source1d, points, search, radii, sink, processes)
ap._finalize_processing(verbose=verbose,
function='measure_max',
timer=timer)
return sink
def smooth_by_counting(source, sink = None, low = 5, high = 10, shape = None):
"""Smooth binary image by counting neighbours.
Arguments
---------
source : array
The binary source to smooth.
sink : array or None.
The sink to write the smoothed source to.
low : int
If a voxel has less then this number of 26-neighbours it is set to False.
high : int
If a voxel has more then this number of 26-neighbours it is set to True.
shape : tuple of int or None
The shape of the square structuring element to consider.
Returns
-------
sink : array
The smoothed sbinary source.
Note
----
The algorithm uses a sequence of 1d convoluions for speed, allowing only
rectangular like structuring elements.
"""
ndim = source.ndim;
if shape is None:
shape = (3,) * ndim;
filtered = source;
for d in range(ndim):
weights = np.ones(shape[d], dtype = int);
temp = np.zeros(source.shape, dtype = 'uint8');
ap.correlate1d(filtered, weights, sink=temp, axis=d, mode='constant', cval=0);
filtered = temp;
if sink is None:
sink = np.array(source, dtype = bool);
sink[filtered >= high] = True;
sink[filtered < low] = False;
return sink
def clip(source, sink = None, clip_min = None, clip_max = None, clip_norm = None, processes = None, verbose = False):
"""Clip and normalize data.
Arguments
---------
source : array
Input source.
sink : array, dtype or None
output sink or output data type, if None, a new array is allocated.
clip_min : number
Minimal number to clip source data to.
clip_max : number
Maximal number to clip source data to.
clip_norm : number
Normalization constant.
Returns
-------
sink : array
Clipped output.
"""
processes, timer = ap.initialize_processing(verbose=verbose, processes=processes, function='clip');
source, source_buffer = ap.initialize_source(source);
if source.ndim != 3:
raise ValueError('Source assumed to be 3d found %dd!' % source.ndim);
if clip_min is None:
clip_min = ap.io.min_value(source);
if clip_max is None:
clip_max = ap.io.max_value(source);
if clip_norm is None:
clip_norm = clip_max - clip_min;
sink, sink_buffer = ap.initialize_sink(sink = sink, source = source);
code.clip(source_buffer, sink_buffer, clip_min, clip_max, clip_norm, processes);
return sink;
def rank(source,
sink=None,
function=rnk.median,
resample=None,
verbose=False,
out=sys.stdout,
**kwargs):
"""Rank filter inbetween reshaping."""
timer = tmr.Timer()
sink, sink_buffer = ap.initialize_sink(sink=sink, source=source, order='F')
if resample:
interpolation = cv2.INTER_NEAREST
new_shape = np.round(np.array(sink.shape, dtype=float) *
resample).astype(int)
new_shape[2] = sink.shape[2]
data = np.zeros(tuple(new_shape), order='F', dtype=source.dtype)
new_shape = tuple(new_shape[1::-1])
for z in range(source.shape[2]):
data[:, :, z] = cv2.resize(src=source[:, :, z],
dsize=new_shape,
interpolation=interpolation)
#print data.shape, data.dtype
out.write(timer.elapsed_time(head='Rank filter: Resampling') + '\n')
else:
data = source
#keys = inspect.getargspec(function).args;
#kwargs = { k : v for k,v in kwargs.iteritems() if k in keys};
data = function(data, **kwargs)
out.write(
timer.elapsed_time(head='Rank filter: %s' % function.__name__) + '\n')
if resample:
#interpolation = cv2.INTER_LINEAR;
interpolation = cv2.INTER_AREA
for z in range(sink.shape[2]):
sink_buffer[:, :, z] = cv2.resize(src=data[:, :, z],
dsize=sink.shape[1::-1],
interpolation=interpolation)
out.write(timer.elapsed_time(head='Rank filter: Upsampling') + '\n')
else:
sink_buffer[:] = data
return sink
def detect_maxima(source, h_max=None, shape=5, threshold=None, verbose=False):
# extended maxima
maxima = md.find_maxima(source,
h_max=h_max,
shape=shape,
threshold=threshold,
verbose=verbose)
#center of maxima
if h_max:
centers = md.find_center_of_maxima(source,
maxima=maxima,
verbose=verbose)
else:
centers = ap.where(maxima).array
return centers
def detect_cells(
source,
sink=None,
cell_detection_parameter=default_cell_detection_parameter,
processing_parameter=default_cell_detection_processing_parameter):
"""Cell detection pipeline.
Arguments
---------
source : source specification
The source of the stitched raw data.
sink : sink specification or None
The sink to write the result to. If None, an array is returned.
cell_detection_parameter : dict
Parameter for the binarization. See below for details.
processing_parameter : dict
Parameter for the parallel processing.
See :func:`ClearMap.ParallelProcessing.BlockProcesing.process` for
description of all the parameter.
verbose : bool
If True, print progress output.
Returns
-------
sink : Source
The result of the cell detection.
Notes
-----
Effectively this function performs the following steps:
* illumination correction via :func:`~ClearMap.ImageProcessing.IlluminationCorrection.correct_illumination`
* background removal
* difference of Gaussians (DoG) filter
* maxima detection via :func:`~ClearMap.Analysis.Measurements.MaximaDetection.find_extended_maxima`
* cell shape detection via :func:`~ClearMap.Analysis.Measurements.ShapeDetection.detect_shape`
* cell intensity and size measurements via: :func:`~ClearMap.ImageProcessing.Measurements.ShapeDetection.find_intensity`,
:func:`~ClearMap.ImageProcessing.Measurements.ShapeDetection.find_size`.
The parameters for each step are passed as sub-dictionaries to the
cell_detection_parameter dictionary.
* If None is passed for one of the steps this step is skipped.
* Each step also has an additional parameter 'save' that enables saving of
the result of that step to a file to inspect the pipeline.
Illumination correction
-----------------------
illumination_correction : dict or None
Illumination correction step parameter.
flatfield : array or str
The flat field estimate for the image planes.
background : array or None
A background level to assume for the flatfield correction.
scaling : float, 'max', 'mean' or None
Optional scaling after the flat field correction.
save : str or None
Save the result of this step to the specified file if not None.
See also :func:`ClearMap.ImageProcessing.IlluminationCorrection.correct_illumination`
Background removal
------------------
background_correction : dict or None
Background removal step parameter.
shape : tuple
The shape of the structure lement to estimate the background.
This should be larger than the typical cell size.
form : str
The form of the structur element (e.g. 'Disk')
save : str or None
Save the result of this step to the specified file if not None.
Equalization
------------
equalization : dict or None
Equalization step parameter.
See also :func:`ClearMap.ImageProcessing.LocalStatistics.local_percentile`
precentile : tuple
The lower and upper percentiles used to estimate the equalization.
The lower percentile is used for normalization, the upper to limit the
maximal boost to a maximal intensity above this percentile.
max_value : float
The maximal intensity value in the equalized image.
selem : tuple
The structural element size to estimate the percentiles.
Should be larger than the larger vessels.
spacing : tuple
The spacing used to move the structural elements.
Larger spacings speed up processing but become locally less precise.
interpolate : int
The order of the interpoltation used in constructing the full
background estimate in case a non-trivial spacing is used.
save : str or None
Save the result of this step to the specified file if not None.
DoG Filter
----------
dog_filter : dict or None
Difference of Gaussian filter step parameter.
shape : tuple
The shape of the filter.
This should be near the typical cell size.
sigma : tuple or None
The std of the inner Gaussian.
If None, detemined automatically from shape.
sigma2 : tuple or None
The std of the outer Gaussian.
If None, detemined automatically from shape.
save : str or None
Save the result of this step to the specified file if not None.
Maxima detection
----------------
maxima_detection : dict or None
Extended maxima detection step parameter.
h_max : float or None
The 'height'for the extended maxima.
If None, simple local maxima detection isused.
shape : tuple
The shape of the structural element for extended maxima detection.
This should be near the typical cell size.
threshold : float or None
Only maxima above this threshold are detected. If None, all maxima
are detected.
valid : bool
If True, only detect cell centers in the valid range of the blocks with
overlap.
save : str or None
Save the result of this step to the specified file if not None.
Shape detection
---------------
shape_detection : dict or None
Shape detection step parameter.
threshold : float
Cell shape is expanded from maxima if pixles are above this threshold
and not closer to another maxima.
save : str or None
Save the result of this step to the specified file if not None.
Intensity detection
-------------------
intensity_detection : dict or None
Intensity detection step parameter.
method : {'max'|'min','mean'|'sum'}
The method to use to measure the intensity of a cell.
shape : tuple or None
If no cell shapes are detected a disk of this shape is used to measure
the cell intensity.
save : str or None
Save the result of this step to the specified file if not None.
References
----------
[1] Renier, Adams, Kirst, Wu et al., "Mapping of Brain Activity by Automated Volume Analysis of Immediate Early Genes.", Cell 165, 1789 (2016)
[1] Kirst et al., "Mapping the Fine-Scale Organization and Plasticity of the Brain Vasculature", Cell 180, 780 (2020)
"""
#initialize sink
shape = io.shape(source)
order = io.order(source)
for key in cell_detection_parameter.keys():
par = cell_detection_parameter[key]
if isinstance(par, dict):
filename = par.get('save', None)
if filename:
ap.initialize_sink(filename,
shape=shape,
order=order,
dtype='float')
cell_detection_parameter.update(
verbose=processing_parameter.get('verbose', False))
results, blocks = bp.process(detect_cells_block,
source,
sink=None,
function_type='block',
return_result=True,
return_blocks=True,
parameter=cell_detection_parameter,
**processing_parameter)
#merge results
results = np.vstack([np.hstack(r) for r in results])
#create column headers
header = ['x', 'y', 'z']
dtypes = [int, int, int]
if cell_detection_parameter['shape_detection'] is not None:
header += ['size']
dtypes += [int]
measures = cell_detection_parameter['intensity_detection']['measure']
header += measures
dtypes += [float] * len(measures)
dt = {
'names': header,
'formats': dtypes
}
cells = np.zeros(len(results), dtype=dt)
for i, h in enumerate(header):
cells[h] = results[:, i]
#save results
return io.write(sink, cells)
def connectPoint(data,
mask,
endpoints,
start_index,
radius,
tubeness=None,
min_quality=None,
remove_local_mask=True,
skeleton=None,
verbose=False,
**trace_parameter):
"""Tries to connect an end point"""
#outine:
# find neighbour end points and try to connect to nearest one
# if path score good enough add path and remove two endpoints
# else try to connect to binarized image
# if path score good enugh connect to closest skeleton point
# else not connectable
#assumes everything is in fotran order
strides = np.array(data.strides) / data.itemsize
shape = data.shape
#print strides, shape
center_flat = endpoints[start_index]
center_xyz = np.array(np.unravel_index(center_flat, data.shape, order='F'))
mask_nbh = extractNeighbourhood(mask, center_xyz, radius)
data_nbh = np.asarray(extractNeighbourhood(data, center_xyz, radius),
dtype=float,
order='F')
shape_nbh = mask_nbh.shape
center_nbh_xyz = np.zeros(3, dtype=int) + radius
#center_nbh_flat = np.ravel_multi_index(center_nbh_xyz, shape_nbh, order = 'F');
if tubeness is None:
tubeness_nbh = cur.tubeness(
ndi.gaussian_filter(np.asarray(data_nbh, dtype=float), sigma=1.0))
tubeness_nbh = np.asarray(tubeness_nbh, order='F')
else:
tubeness_nbh = extractNeighbourhood(tubeness, center_xyz, radius)
mask_nbh_label = np.empty(shape_nbh, dtype='int32', order='F')
_ = ndi.label(mask_nbh,
structure=np.ones((3, 3, 3), dtype=bool),
output=mask_nbh_label)
local_nbh = mask_nbh_label[tuple(center_nbh_xyz)] == mask_nbh_label
# end point neighbours
nbs_flat = ap.findNeighbours(endpoints, start_index, shape, strides,
radius)
if len(nbs_flat) > 0:
nbs_nbh_xyz = np.vstack(np.unravel_index(
nbs_flat, shape, order='F')).T - center_xyz + center_nbh_xyz
nbs_nbh_flat = np.ravel_multi_index(nbs_nbh_xyz.T,
shape_nbh,
order='F')
# remove connected neighbours
non_local_nbh_flat = np.reshape(np.logical_not(local_nbh),
-1,
order='F')
nbs_nbh_non_local_flat = nbs_nbh_flat[non_local_nbh_flat[nbs_nbh_flat]]
if len(nbs_nbh_non_local_flat) > 0:
#find nearest neighbour
nbs_nbh_non_local_xyz = np.vstack(
np.unravel_index(nbs_nbh_non_local_flat, shape, order='F')).T
nbs_nbh_non_local_dist = nbs_nbh_non_local_xyz - center_nbh_xyz
nbs_nbh_non_local_dist = np.sum(nbs_nbh_non_local_dist *
nbs_nbh_non_local_dist,
axis=1)
neighbor_nbh_xyz = nbs_nbh_non_local_xyz[np.argmin(
nbs_nbh_non_local_dist)]
path, quality = trc.trace(data_nbh,
tubeness_nbh,
center_nbh_xyz,
neighbor_nbh_xyz,
verbose=False,
returnQuality=True,
**trace_parameter)
if len(path) > 0:
if quality / len(path) < min_quality:
if verbose:
print(
'Found good path to neighbour of length = %d with quality = %f (per length = %f) [%d / %d nonlocal neighbours]'
% (len(path), quality, quality / len(path),
len(nbs_nbh_non_local_flat), len(nbs_flat)))
#print path
return path + center_xyz - center_nbh_xyz, quality
else:
if verbose:
print(
'Found bad path to neighbour of length = %d with quality = %f (per length = %f) [%d / %d nonlocal neighbours]'
% (len(path), quality, quality / len(path),
len(nbs_nbh_non_local_flat), len(nbs_flat)))
#print path
else:
if verbose:
print(
'Found no path to neighbour [%d / %d nonlocal neighbours]'
% (len(nbs_nbh_non_local_flat), len(nbs_flat)))
#print path
#tracing to neares neighbour failed
if verbose:
print('Found no valid path to neighbour, now tracing to binary!')
#print path
# Tracing to next binary
if remove_local_mask:
mask_nbh[local_nbh] = False
distance_nbh = ndi.distance_transform_edt(np.logical_not(mask_nbh))
distance_nbh = np.asarray(distance_nbh, order='F')
path, quality = trc.traceToMask(data_nbh,
tubeness_nbh,
center_nbh_xyz,
distance_nbh,
verbose=False,
returnQuality=True,
**trace_parameter)
if len(path) > 0:
if quality / len(path) < min_quality:
if verbose:
print(
'Found good path to binary of length = %d with quality = %f (per length = %f)'
% (len(path), quality, quality / len(path)))
#print path
# trace to skeleton
if skeleton is not None:
#find closest point on skeleton
final_xyz = path[0]
skeleton_nbh = extractNeighbourhood(skeleton, center_xyz,
radius)
local_end_path_nbh = mask_nbh_label[tuple(
final_xyz)] == mask_nbh_label
skeleton_nbh_dxyz = np.vstack(
np.where(np.logical_and(skeleton_nbh,
local_end_path_nbh))).T - final_xyz
if len(
skeleton_nbh_dxyz
) == 0: # could not find skeleton nearby -> give up for now
return path + center_xyz - center_nbh_xyz, quality
skeleton_nbh_dist = np.sum(skeleton_nbh_dxyz *
skeleton_nbh_dxyz,
axis=1)
closest_dxyz = skeleton_nbh_dxyz[np.argmin(skeleton_nbh_dist)]
closest_xyz = closest_dxyz + final_xyz
#print path[0], path[-1]
#print center_nbh_xyz, closest_dxyz
#generate pixel path
max_l = np.max(np.abs(closest_dxyz)) + 1
path_add_xyz = np.vstack([
np.asarray(np.linspace(f, c, max_l), dtype=int)
for f, c in zip(final_xyz, closest_xyz)
]).T
path_add_flat = np.ravel_multi_index(path_add_xyz.T, shape_nbh)
_, ids = np.unique(path_add_flat, return_index=True)
path_add_xyz = path_add_xyz[ids]
#print path_add_xyz;
path = np.vstack([path, path_add_xyz])
# note: this is not an ordered path anymore!
return path + center_xyz - center_nbh_xyz, quality
else:
if verbose:
print(
'Found bad path to binary of length = %d with quality = %f (per length = %f)'
% (len(path), quality, quality / len(path)))
#print path
if verbose:
print('Found no valid path to binary!')
return np.zeros((0, 3)), 0
def average(source,
sink=None,
shape=None,
dtype=None,
weights=None,
indices=None,
kernel=None,
return_counts=False,
processes=None,
verbose=False):
"""Averages a list of points into an volumetric image array.
Arguments
---------
source : str, array or Source
Source of point of nxd coordinates.
sink : str, array or None
The sink for the devolved image, if None return array.
shape : tuple, str or None
Shape of the final devolved data. If None, determine from points.
If str, determine shape from the source at the specified location.
dtype : dtype or None
Optional data type of the sink.
weights : array or None
Weight array of length n for each point. If None, use uniform weights.
method : str
Method for voxelization: 'sphere', 'rectangle' or 'pixel'.
indices : array
The relative indices to the center to devolve over as nxd array.
kernel : array
Optional kernel weights for each index in indices.
processes : int or None
Number of processes to use.
verbose : bool
If True, print progress info.
Returns
-------
sink : str, array
Volumetric data of devolved point data.
"""
processes, timer = ap.initialize_processing(processes=processes,
verbose=verbose,
function='devolve')
#points, points_buffer = ap.initialize_source(points);
points_buffer = io.read(source)
if points_buffer.ndim == 1:
points_buffer = points_buffer[:, None]
if sink is None and shape is None:
if points_buffer.ndim > 1:
shape = tuple(
int(math.ceil(points_buffer[:, d].max()))
for d in range(points_buffer.shape[1]))
else:
shape = (int(math.ceil(points_buffer[:].max())), )
elif isinstance(shape, str):
shape = io.shape(shape)
if sink is None and dtype is None:
if weights is not None:
dtype = io.dtype(weights)
elif kernel is not None:
kernel = np.asarray(kernel)
dtype = kernel.dtype
else:
dtype = int
sink, sink_buffer, sink_shape, sink_strides = ap.initialize_sink(
sink=sink,
shape=shape,
dtype=dtype,
return_shape=True,
return_strides=True,
as_1d=True)
#TODO: initialize properly
counts = np.zeros(sink_shape, dtype=int, order=sink.order)
counts_buffer = counts.reshape(-1, order='A')
#print(counts.shape, counts_buffer.shape)
if indices is None:
return sink
indices = np.asarray(indices, dtype=int)
if indices.ndim == 1:
indices = indices[:, None]
if kernel is not None:
kernel = np.asarray(kernel, dtype=float)
#print(kernel);
#print(weights)
#return;
code.average(points_buffer, weights, indices, sink_buffer, sink_shape,
sink_strides, counts_buffer, processes)
# if weights is None:
# if kernel is None:
# code.devolve_uniform(points_buffer, indices, sink_buffer, sink_shape, sink_strides, processes);
# else:
# code.devolve_uniform_kernel(points_buffer, indices, kernel, sink_buffer, sink_shape, sink_strides, processes);
# else:
# if kernel is None:
# code.devolve_weights(points_buffer, weights, indices, sink_buffer, sink_shape, sink_strides, processes);
# else:
# code.devolve_weights_kernel(points_buffer, weights, indices, kernel, sink_buffer, sink_shape, sink_strides, processes);
#TODO: move to code
good = counts_buffer > 0
sink_buffer[good] /= counts_buffer[good]
ap.finalize_processing(verbose=verbose, function='devolve', timer=timer)
if return_counts:
return sink, counts
else:
return sink
def binarize(source, sink = None, binarization_parameter = default_binarization_parameter, processing_parameter = default_binarization_processing_parameter):
"""Multi-path binarization of iDISCO+ cleared vasculature data.
Arguments
---------
source : source specification
The source of the stitched raw data.
sink : sink specification or None
The sink to write the result to. If None, an array is returned.
binarization_parameter : dict
Parameter for the binarization. See below for details.
processing_parameter : dict
Parameter for the parallel processing.
See :func:`ClearMap.ParallelProcessing.BlockProcesing.process` for
description of all the parameter.
verbose : bool
If True, print progress output.
Returns
-------
sink : Source
The result of the binarization.
Notes
-----
* The binarization pipeline is composed of several steps. The parameters for
each step are passed as sub-dictionaries to the binarization_parameter
dictionary.
* If None is passed for one of the steps this step is skipped.
* Each step also has an additional parameter 'save' that enables saving of
the result of that step to a file to inspect the pipeline.
General parameter
-----------------
binary_status : str or None
File name to save the information about which part of the multi-path
binarization contributed to the final result.
max_bin : int
Number of intensity levels to use for the data after preprocessing.
Higher values will increase the intensity resolution but slow down
processing.
For the vasculature a typical value is 2**12.
Clipping
--------
clip : dict or None
Clipping and mask generation step parameter.
clip_range : tuple
The range to clip the raw data as (lowest, highest)
Voxels above lowest define the foregournd mask used
in the following steps.
For the vasculature a typical value is (400,60000).
save : str or None
Save the result of this step to the specified file if not None.
See also :mod:`ClearMap.ImageProcessing.Clipping.Clipping`
Lightsheet correction
---------------------
lightsheet : dict or None
Lightsheet correction step parameter.
percentile : float
Percentile in [0,1] used to estimate the lightshieet artifact.
For the vasculature a typical value is 0.25.
lightsheet : dict
Parameter for the ligthsheet artifact percentile estimation.
See :func:`ClearMap.ImageProcessing.LightsheetCorrection.correct_lightsheet`
for list of all parameters. The crucial parameter is
selem : tuple
The structural element shape used to estimate the stripe artifact.
It should match the typical lenght, width, and depth of the artifact
in the data.
For the vasculature a typical value is (150,1,1).
background : dict
Parameter for the background estimation in the light sheet correction.
See :func:`ClearMap.ImageProcessing.LightsheetCorrection.correct_lightsheet`
for list of all parameters. The crucial parameters are
selem : tuple
The structural element shape used to estimate the background.
It should be bigger than the largest vessels,
For the vasculature a typical value is (200,200,1).
spacing : tuple
The spacing to use to estimate the background. Larger spacings speed up
processing but become less local estimates.
For the vasculature a typical value is (25,25,1)
step : tuple
This parameter enables to subsample from the entire array defined by
the structural element using larger than single voxel steps.
For the vasculature a typical value is (2,2,1).
interpolate : int
The order of the interpoltation used in constructing the full
background estimate in case a non-trivial spacing is used.
For the vasculature a typical value is 1.
lightsheet_vs_background : float
The background is multiplied by this weight before comparing to the
lightsheet artifact estimate.
For the vasculature a typical value is 2.
save : str or None
Save the result of this step to the specified file if not None.
Median filter
-------------
median : dict or None
Median correction step parameter.
See :func:`ClearMap.ImageProcessing.Filter.Rank.median` for all parameter.
The important parameters are
selem : tuple
The structural element size for the median filter.
For the vascualture a typical value is (3,3,3).
save : str or None
Save the result of this step to the specified file if not None.
Pseudo Deconvolution
--------------------
deconvolve : dict
The deconvolution step parameter.
sigma : float
The std of a Gaussina filter applied to the high intensity pixel image.
The number should reflect the scale of the halo effect seen around high
intensity structures.
For the vasculature a typical value is 10.
save : str or None
Save the result of this step to the specified file if not None.
threshold : float
Voxels above this threshold will be added to the binarization result
in the multi-path biniarization.
For the vasculature a typical value is 750.
Adaptive Thresholding
---------------------
adaptive : dict or None
Adaptive thresholding step parameter.
A local ISODATA threshold is estimated.
See also :mod:`ClearMap.ImageProcessing.LocalStatistics`.
selem : tuple
The structural element size to estimate the percentiles.
Should be larger than the larger vessels.
For the vasculature a typical value is (200,200,5).
spacing : tuple
The spacing used to move the structural elements.
Larger spacings speed up processing but become locally less precise.
For the vasculature a typical value is (50,50,5)
interpolate : int
The order of the interpoltation used in constructing the full
background estimate in case a non-trivial spacing is used.
For the vasculature a typical value is 1.
save : str or None
Save the result of this step to the specified file if not None.
Equalization
------------
equalize : dict or None
Equalization step parameter.
See also :func:`ClearMap.ImageProcessing.LocalStatistics.local_percentile`
precentile : tuple
The lower and upper percentiles used to estimate the equalization.
The lower percentile is used for normalization, the upper to limit the
maximal boost to a maximal intensity above this percentile.
For the vasculature a typical value is (0.4, 0.975).
max_value : float
The maximal intensity value in the equalized image.
For the vasculature a typical value is 1.5.
selem : tuple
The structural element size to estimate the percentiles.
Should be larger than the larger vessels.
For the vasculature a typical value is (200,200,5).
spacing : tuple
The spacing used to move the structural elements.
Larger spacings speed up processing but become locally less precise.
For the vasculature a typical value is (50,50,5)
interpolate : int
The order of the interpoltation used in constructing the full
background estimate in case a non-trivial spacing is used.
For the vasculature a typical value is 1.
save : str or None
Save the result of this step to the specified file if not None.
threshold : float
Voxels above this threshold will be added to the binarization result
in the multi-path biniarization.
For the vasculature a typical value is 1.1.
Tube filter
-----------
vesselize : dict
The tube filter step parameter.
background : dict or None
Parameters to correct for local background. See
:func:`ClearMap.ImageProcessing.Filter.Rank.percentile`.
If None, no background correction is done before the tube filter.
selem : tuple
The structural element specification to estimate the percentiles.
Should be larger than the largest vessels intended to be
boosted by the tube filter.
For the vasculature a typical value is ('disk', (30,30,1)) .
percentile : float
Percentile in [0,1] used to estimate the background.
For the vasculature a typical value is 0.5.
tubness : dict
Parameters used for the tube filter. See
:func:`ClearMap.ImageProcessing.Differentiation.Hessian.lambda123`.
sigma : float
The scale of the vessels to boos in the filter.
For the vasculature a typical value is 1.0.
save : str or None
Save the result of this step to the specified file if not None.
threshold : float
Voxels above this threshold will be added to the binarization result
in the multi-path biniarization.
For the vasculature a typical value is 120.
Binary filling
--------------
fill : dict or None
If not None, apply a binary filling the binarized result.
For the vasculature this step is set to None and done globally
in the postprocessing step.
Binary smoothing
----------------
smooth : dict or None
The smoothing parameter passed to
:func:`ClearMap.ImageProcessing.Binary.Smoothing.smooth_by_configuration`.
For the vasculature this step is set to None and done globally
in the postprocessing step.
References
----------
[1] C. Kirst et al., "Mapping the Fine-Scale Organization and Plasticity of the Brain Vasculature", Cell 180, 780 (2020)
"""
#initialize sink
shape = io.shape(source);
order = io.order(source);
sink, sink_buffer = ap.initialize_sink(sink=sink, shape=shape, order=order, dtype=bool); #, memory='shared');
#initialize addition output sinks
binary_status = binarization_parameter.get('binary_status', None);
if binary_status:
ap.initialize_sink(binary_status, source=sink, shape=shape, order=order, dtype='uint16');
for key in binarization_parameter.keys():
par = binarization_parameter[key];
if isinstance(par, dict):
filename = par.get('save', None);
if filename:
ap.initialize_sink(filename, shape=shape, order=order, dtype='float');
binarization_parameter.update(verbose=processing_parameter.get('verbose', False));
bp.process(binarize_block, source, sink, function_type='block', parameter=binarization_parameter, **processing_parameter)
return sink;
def postprocess(source, sink = None, postprocessing_parameter = default_postprocessing_parameter, processing_parameter = default_postprocessing_processing_parameter, processes = None, verbose = True):
"""Postprocess a binarized image.
Arguments
---------
source : source specification
The binary source.
sink : sink specification or None
The sink to write the postprocesses result to.
If None, an array is returned.
postprocessing_parameter : dict
Parameter for the postprocessing.
processing_parameter : dict
Parameter for the parallel processing.
verbose : bool
If True, print progress output.
Returns
-------
sink : Source
The result of the binarization.
Notes
-----
* The postporcessing pipeline is composed of several steps. The parameters
for each step are passed as sub-dictionaries to the
postprocessing_parameter dictionary.
* If None is passed for one of the steps the step is skipped.
Smoothing
---------
smooth : dict or None
Smoothing step parameter. See
:func:`ClearMap.ImageProcessing.Binary.Smoothing.smooth_by_configuration`
iterations : int
Number of smoothing iterations.
For the vasculature a typical value is 6.
Filling
-------
fill : bool or None
If True, fill holes in the binary data.
"""
source = io.as_source(source);
sink = ap.initialize_sink(sink, shape=source.shape, dtype=source.dtype, order=source.order, return_buffer=False);
if verbose:
timer = tmr.Timer();
print('Binary post processing: initialized.');
postprocessing_parameter = postprocessing_parameter.copy();
parameter_smooth = postprocessing_parameter.pop('smooth', None);
parameter_fill = postprocessing_parameter.pop('fill', None);
#print(parameter_smooth, parameter_fill)
#smoothing
save = None;
if parameter_smooth:
#intialize temporary files if needed
if parameter_fill:
save = parameter_smooth.pop('save', None);
temporary_filename = save;
if temporary_filename is None:
temporary_filename = postprocessing_parameter['temporary_filename'];
if temporary_filename is None:
temporary_filename = tmpf.mktemp(prefix='TubeMap_Vasculature_postprocessing', suffix='.npy');
sink_smooth = ap.initialize_sink(temporary_filename, shape=source.shape, dtype=source.dtype, order=source.order, return_buffer=False);
else:
sink_smooth = sink;
#run smoothing
source_fill = bs.smooth_by_configuration(source, sink=sink_smooth, processing_parameter=processing_parameter, processes=processes, verbose=verbose, **parameter_smooth);
else:
source_fill = source;
if parameter_fill:
sink = bf.fill(source_fill, sink=sink, processes=processes, verbose=verbose);
if parameter_smooth and save is None:
io.delete_file(temporary_filename);
else:
sink = source_fill;
if verbose:
timer.print_elapsed_time('Binary post processing');
gc.collect()
return None;
def _test():
import numpy as np
import ClearMap.ParallelProcessing.DataProcessing.ArrayProcessing as ap
from importlib import reload
reload(ap)
## Lookup table processing
#apply_lut
x = np.random.randint(0, 100, size=(20,30));
lut = np.arange(100) + 1;
y = ap.apply_lut(x, lut)
assert np.all(y == x+1)
#apply_lut_to_index
import ClearMap.ImageProcessing.Topology.Topology3d as t3d
kernel = t3d.index_kernel(dtype=int);
import ClearMap.ImageProcessing.Binary.Smoothing as sm
lut = sm.initialize_lookup_table()
data = np.random.randint(0, 2, (1500,300,400), dtype = bool)
#reload(ap)
result = ap.apply_lut_to_index(data, kernel, lut, sink=None, verbose=True)
import ClearMap.Visualization.Plot3d as p3d
p3d.plot([data, result])
## Correlation
#correlate1d
#reload(ap)
axis = 1;
kernel = np.array(range(11), dtype='uint32');
data = np.random.randint(0, 2**27, (1000, 1500,100), dtype='uint32');
corr = ap.correlate1d(data, kernel, axis=axis, verbose=True, processes=10);
import scipy.ndimage as ndi
import ClearMap.Utils.Timer as tmr
timer = tmr.Timer();
corr_ndi = ndi.correlate1d(data, kernel, axis=axis, mode='constant',cval=0);
timer.print_elapsed_time('ndi')
assert np.allclose(corr, corr_ndi)
#
#
#
#
#
#default_blocks_per_process = 10;
#"""Default number of blocks per process to split the data.
#
#Note
#----
#10 blocks per process is a good choice.
#"""
#
#default_cutoff = 20000000;
#"""Default size of array below which ordinary numpy is used.
#
#Note
#----
#Ideally test this on your machine for different array sizes.
#"""
#
#
#
#def blockRanges(data, blocks = None, processes = defaultProcesses):
# """Ranges of evenly spaced blocks in array
#
# Arguments:
# data : array
# array to divide in blocks
# blocks : int or None
# number of blocks to split array into
# processes : None or int
# number of processes, if None use number of cpus
#
# Returns:
# array
# list of the range boundaries
# """
# if processes is None:
# processes = defaultProcesses;
# if blocks is None:
# blocks = processes * defaultBlocksPerProcess;
#
# d = data.reshape(-1, order = 'A');
# blocks = min(blocks, d.shape[0]);
# return np.array(np.linspace(0, d.shape[0], blocks + 1), dtype = int);
#
#
#def blockSums(data, blocks = None, processes = defaultProcesses):
# """Sums of evenly spaced blocks in array
#
# Arguments:
# data : array
# array to perform the block sums on
# blocks : int or None
# number of blocks to split array into
# processes : None or int
# number of processes, if None use number of cpus
#
# Returns:
# array
# sums of the values in the different blocks
# """
# if processes is None:
# processes = defaultProcesses;
# if blocks is None:
# blocks = processes * defaultBlocksPerProcess;
#
# d = data.reshape(-1, order = 'A');
# if data.dtype == bool:
# d = d.view('uint8')
#
# return code.blockSums1d(d, blocks = blocks, processes = processes);
#
#
#def where(data, out = None, blocks = None, cutoff = defaultCutoff, processes = defaultProcesses):
# """Returns the indices of the non-zero entries of the array
#
# Arguments:
# data : array
# array to search for nonzero indices
# out : array or None
# if not None results is written into this array
# blocks : int or None
# number of blocks to split array into for parallel processing
# cutoff : int
# number of elements below whih to switch to numpy.where
# processes : None or int
# number of processes, if None use number of cpus
#
# Returns:
# array
# positions of the nonzero entries of the input array
#
# Note:
# Uses numpy.where if there is no match of dimension implemented!
# """
# if data.ndim != 1 and data.ndim != 3:
# raise Warning('Using numpy where for dimension %d and type %s!' % (data.ndim, data.dtype))
# return np.vstack(np.where(data)).T;
#
# if cutoff is None:
# cutoff = 1;
# cutoff = min(1, cutoff);
# if data.size <= cutoff:
# return np.vstack(np.where(data)).T;
#
# if processes is None:
# processes = defaultProcesses;
# if blocks is None:
# blocks = processes * defaultBlocksPerProcess;
#
# if data.dtype == bool:
# d = data.view('uint8')
# else:
# d = data;
#
# if out is None:
# if d.ndim == 1:
# sums = code.blockSums1d(d, blocks = blocks, processes = processes);
# else:
# sums = code.blockSums3d(d, blocks = blocks, processes = processes);
# out = np.squeeze(np.zeros((np.sum(sums), data.ndim), dtype = np.int));
# else:
# sums = None;
#
# if d.ndim == 1:
# code.where1d(d, out = out, sums = sums, blocks = blocks, processes = processes);
# else: # d.ndim == 3:
# code.where3d(d, out = out, sums = sums, blocks = blocks, processes = processes);
#
# return out;
#
#
#
#
#def setValue(data, indices, value, cutoff = defaultCutoff, processes = defaultProcesses):
# """Set value at specified indices of an array
#
# Arguments:
# data : array
# array to search for nonzero indices
# indices : array or None
# list of indices to set
# value : numeric or bool
# value to set elements in data to
# processes : None or int
# number of processes, if None use number of cpus
#
# Returns:
# array
# array with specified entries set to new value
#
# Note:
# Uses numpy if there is no match of dimension implemented!
# """
# if data.ndim != 1:
# raise Warning('Using numpy where for dimension %d and type %s!' % (data.ndim, data.dtype))
# data[indices] = value;
# return data;
#
# if cutoff is None:
# cutoff = 1;
# cutoff = min(1, cutoff);
# if data.size <= cutoff:
# data[indices] = value;
# return data;
#
# if processes is None:
# processes = defaultProcesses;
#
# if data.dtype == bool:
# d = data.view('uint8')
# else:
# d = data;
#
# code.set1d(d, indices, value, processes = processes);
#
# return data;
#
#
#def setArray(data, indices, values, cutoff = defaultCutoff, processes = defaultProcesses):
# """Set value at specified indices of an array
#
# Arguments:
# data : array
# array to search for nonzero indices
# indices : array or None
# list of indices to set
# values : array
# values to set elements in data to
# processes : None or int
# number of processes, if None use number of cpus
#
# Returns:
# array
# array with specified entries set to new value
#
# Note:
# Uses numpy if there is no match of dimension implemented!
# """
# if data.ndim != 1:
# raise Warning('Using numpy where for dimension %d and type %s!' % (data.ndim, data.dtype))
# data[indices] = values;
# return data;
#
# if cutoff is None:
# cutoff = 1;
# cutoff = min(1, cutoff);
# if data.size <= cutoff:
# data[indices] = values;
# return data;
#
# if processes is None:
# processes = defaultProcesses;
#
# if data.dtype == bool:
# d = data.view('uint8')
# else:
# d = data;
#
# code.set1darray(d, indices, values, processes = processes);
#
# return data;
#
#
#
#def take(data, indices, out = None, cutoff = defaultCutoff, processes = defaultProcesses):
# """Extracts the values at specified indices
#
# Arguments:
# data : array
# array to search for nonzero indices
# out : array or None
# if not None results is written into this array
# cutoff : int
# number of elements below whih to switch to numpy.where
# processes : None or int
# number of processes, if None use number of cpus
#
# Returns:
# array
# positions of the nonzero entries of the input array
#
# Note:
# Uses numpy data[indices] if there is no match of dimension implemented!
# """
# if data.ndim != 1:
# raise Warning('Using numpy where for dimension %d and type %s!' % (data.ndim, data.dtype))
# return data[indices];
#
# if cutoff is None:
# cutoff = 1;
# cutoff = min(1, cutoff);
# if data.size < cutoff:
# return data[indices];
#
# if processes is None:
# processes = defaultProcesses;
#
# if data.dtype == bool:
# d = data.view('uint8')
# else:
# d = data;
#
# if out is None:
# out = np.empty(len(indices), dtype = data.dtype);
# if out.dtype == bool:
# o = out.view('uint8');
# else:
# o = out;
#
# code.take1d(d, indices, o, processes = processes);
#
# return out;
#
#
#def match(match, indices, out = None):
# """Matches a sorted list of 1d indices to another larger one
#
# Arguments:
# match : array
# array of indices to match to indices
# indices : array or None
# array of indices
#
# Returns:
# array
# array with specified entries set to new value
#
# Note:
# Uses numpy if there is no match of dimension implemented!
# """
# if match.ndim != 1:
# raise ValueError('Match array dimension required to be 1d, found %d!' % (match.ndim))
# if indices.ndim != 1:
# raise ValueError('Indices array dimension required to be 1d, found %d!' % (indices.ndim))
#
# if out is None:
# out = np.empty(len(match), dtype = match.dtype);
#
# code.match1d(match, indices, out);
#
# return out;
#
#
# Find neighbours in an index list
#
#
#def neighbours(indices, offset, processes = defaultProcesses):
# """Returns all pairs of indices that are apart a specified offset"""
# return code.neighbours(indices, offset = offset, processes = processes);
#
#
#def findNeighbours(indices, center, shape, strides, mask):
# """Finds all indices within a specified kernel region centered at a point"""
#
# if len(strides) != 3 or len(shape) != 3 or (strides[0] != 1 and strides[2] != 1):
# raise RuntimeError('only 3d C or F contiguous arrays suported');
#
# if isinstance(mask, int):
# mask = (mask,);
# if isinstance(mask, tuple):
# mask = mask * 3;
# return code.neighbourlistRadius(indices, center, shape[0], shape[1], shape[2],
# strides[0], strides[1], strides[2],
# mask[0], mask[1], mask[2]);
# else:
# if mask.dtype == bool:
# mask = mask.view(dtype = 'uint8');
#
# return code.neighbourlistMask(indices, center, shape[0], shape[1], shape[2], strides[0], strides[1], strides[2], mask);
#
# Loading and saving
#
#def readNumpyHeader(filename):
# """Read numpy array information including offset to data
#
# Arguments:
# filename : str
# file name of the numpy file
#
# Returns:
# shape : tuple
# shape of the array
# dtype : dtype
# data type of array
# order : str
# 'C' for c and 'F' for fortran order
# offset : int
# offset in bytes to data buffer in file
# """
# with open(filename, 'rb') as fhandle:
# major, minor = np.lib.format.read_magic(fhandle);
# shape, fortran, dtype = np.lib.format.read_array_header_1_0(fhandle);
# offset = fhandle.tell()
#
# order = 'C';
# if fortran:
# order = 'F';
#
# return (shape, dtype, order, offset)
#
#
#def _offsetFromSlice(sourceSlice, order = 'F'):
# """Checks if slice is compatible with the large data loader and returns z coordiante"""
#
# if order == 'C':
# os = 1; oe = 3; oi = 0;
# else:
# os = 0; oe = 2; oi = 2;
#
# for s in sourceSlice[os:oe]:
# if s.start is not None or s.stop is not None or s.step is not None:
# raise RuntimeError('sub-regions other than in slowest dimension %d not supported! slice = %r' % (oi, sourceSlice))
#
# s = sourceSlice[oi];
# if s.step is not None:
# raise RuntimeError('sub-regions with non unity steps not supported')
#
# if s.start is None:
# s = 0;
# else:
# s = s.start;
#
# return s;
#
#
#def load(filename, region = None, shared = False, blocks = None, processes = cpu_count(), verbose = False):
# """Load a large npy array into memory in parallel
#
# Arguments:
# filename : str
# filename of array to load
# region : Region or None
# if not None this specifies the sub-region to read
# shared : bool
# if True read into shared memory
# blocks : int or None
# number of blocks to split array into for parallel processing
# processes : None or int
# number of processes, if None use number of cpus
# verbose : bool
# print info about the file to be loaded
#
# Returns:
# array
# the data as numpy array
# """
# if processes is None:
# processes = cpu_count();
# if blocks is None:
# blocks = processes * defaultBlocksPerProcess;
#
# #get specs from header specs
# shape, dtype, order, offset = readNumpyHeader(filename);
# if verbose:
# timer = tmr.Timer();
# print('Loading array of shape = %r, dtype = %r, order = %r, offset = %r' %(shape, dtype, order, offset));
#
# if region is not None:
# shape = region.shape();
# sourceSlice = region.sourceSlice();
# off = _offsetFromSlice(sourceSlice, order = order);
#
# if shared:
# data = shm.create(shape, dtype = dtype, order = order);
# else:
# data = np.empty(shape, dtype = dtype, order = order);
#
# d = data.reshape(-1, order = 'A');
# if dtype == bool:
# d = d.view('uint8');
#
# if region is not None:
# if order == 'F':
# offset += data.strides[-1] * off;
# else:
# offset += data.strides[1] * off;
#
# code.load(data = d, filename = filename, offset = offset, blocks = blocks, processes = processes);
#
# if verbose:
# timer.printElapsedTime(head = 'Loading array from %s' % filename);
#
# return data;
#
#
#
#
#def save(filename, data, region = None, blocks = None, processes = cpu_count(), verbose = False):
# """Save a large npy array to disk in parallel
#
# Arguments:
# filename : str
# filename of array to load
# data : array
# array to save to disk
# blocks : int or None
# number of blocks to split array into for parallel processing
# processes : None or int
# number of processes, if None use number of cpus
# verbose : bool
# print info about the file to be loaded
#
# Returns:
# str
# the filename of the numpy array on disk
# """
# if processes is None:
# processes = cpu_count();
# if blocks is None:
# blocks = processes * defaultBlocksPerProcess;
#
# if region is None:
# #create file on disk via memmap
# memmap = np.lib.format.open_memmap(filename, mode = 'w+', shape = data.shape, dtype = data.dtype, fortran_order = np.isfortran(data));
# memmap.flush();
# del(memmap);
#
# #get specs from header specs
# shape, dtype, order, offset = readNumpyHeader(filename);
# if verbose:
# timer = tmr.Timer();
# print('Saving array of shape = %r, dtype = %r, order = %r, offset = %r' %(shape, dtype, order, offset));
#
# if (np.isfortran(data) and order != 'F') or (not np.isfortran(data) and order != 'C'):
# raise RuntimeError('Order of arrays do not match isfortran=%r and order=%s' % (np.isfortran(data), order));
#
# d = data.reshape(-1, order = 'A');
# if dtype == bool:
# d = d.view('uint8');
#
# if region is not None:
# sourceSlice = region.sourceSlice();
# off = _offsetFromSlice(sourceSlice, order = order);
# if order == 'F':
# offset += data.strides[-1] * off;
# else:
# offset += data.strides[1] * off;
#
# #print d.dtype, filename, offset, blocks, processes
#
# code.save(data = d, filename = filename, offset = offset, blocks = blocks, processes = processes);
#
# if verbose:
# timer.printElapsedTime(head = 'Saving array to %s' % filename);
#
# return filename;
#
#
#
#
#
#
#if __name__ == "__main__":
#
# import numpy as np
# from ClearMap.Utils.Timer import Timer;
# import ClearMap.DataProcessing.LargeData as ld
# reload(ld)
#
#
# #dat = np.random.rand(2000,2000,1000) > 0.5;
# #dat = np.random.rand(1000,1000,500) > 0.5;
# dat = np.random.rand(200,300,400) > 0.5;
# #datan = io.MMP.writeData('test.npy', dat);
#
# dat = np.load('data.npy')
# xyz1 = np.load('points.npy')
#
# s = ld.sum(dat)
# print(s == np.sum(s))
#
#
# timer = Timer();
# xyz = ld.where(dat)
# timer.printElapsedTime('parallel')
# #parallel: elapsed time: 0:00:25.807
#
# timer = Timer();
# xyz1 = np.vstack(np.where(dat)).T
# timer.printElapsedTime('numpy')
# #numpy: elapsed time: 0:05:45.590
#
#
# d0 = np.zeros(dat.shape, dtype = bool);
# d1 = np.zeros(dat.shape, dtype = bool);
#
# d0[xyz[:,0], xyz[:,1], xyz[:,2]] = True;
# d1[xyz1[:,0], xyz1[:,1], xyz1[:,2]] = True;
# np.all(d0 == d1)
#
# dat2 = np.array(np.random.rand(1000, 1000, 1000) > 0, dtype = 'bool');
# filename = 'test.npy';
# np.save(filename, dat2)
#
# filename = '/disque/raid/vasculature/4X-test2/170824_IgG_2/170824_IgG_16-23-46/rank_threshold.npy'
#
# timer = Timer();
# ldat = ld.load(filename, verbose = True);
# timer.printElapsedTime('load')
# #load: elapsed time: 0:00:04.867
#
# timer = Timer();
# ldat2 = np.load(filename);
# timer.printElapsedTime('numpy')
# #numpy: elapsed time: 0:00:27.982
#
# np.all(ldat == ldat2)
#
# timer = Timer();
# xyz = ld.where(ldat)
# timer.printElapsedTime('parallel')
# #parallel: elapsed time: 0:07:25.698
#
# lldat = ldat.reshape(-1, order = 'A')
# timer = Timer();
# xyz = ld.where(lldat)
# timer.printElapsedTime('parallel 1d')
# #parallel 1d: elapsed time: 0:00:49.034
#
# timer = Timer();
# xyz = np.where(ldat)
# timer.printElapsedTime('numpy')
#
#
# import os
# #os.remove(filename)
#
# filename = './ClearMap/Test/Skeletonization/test_bin.npy';
# timer = Timer();
# ldat = ld.load(filename, shared = True, verbose = True);
# timer.printElapsedTime('load')
#
# ld.shm.isShared(ldat);
#
#
#
# import numpy as np
# from ClearMap.Utils.Timer import Timer;
# import ClearMap.DataProcessing.LargeData as ld
# reload(ld)
#
# filename = 'test_save.npy';
#
# dat = np.random.rand(100,200,100);
#
# ld.save(filename, dat)
#
#
# dat2 = ld.load(filename)
#
# np.all(dat == dat2)
#
# os.remove(filename)
#
#
#
# import numpy as np
# from ClearMap.Utils.Timer import Timer;
# import ClearMap.DataProcessing.LargeData as ld
# reload(ld)
#
# dat = np.zeros(100, dtype = bool);
# dat2 = dat.copy();
#
# indices = np.array([5,6,7,8,13,42])
#
# ld.setValue(dat, indices, True, cutoff = 0);
#
# dat2[indices] = True;
# np.all(dat2 == dat)
#
# d = ld.take(dat, indices, cutoff = 0)
# np.all(d)
#
#
# import numpy as np
# from ClearMap.Utils.Timer import Timer;
# import ClearMap.DataProcessing.LargeData as ld
# reload(ld)
#
#
# pts = np.array([0,1,5,6,10,11], dtype = int);
#
# ld.neighbours(pts, -10)
#
#
# import numpy as np
# from ClearMap.Utils.Timer import Timer;
# import ClearMap.DataProcessing.LargeData as ld
# import ClearMap.ImageProcessing.Filter.StructureElement as sel;
# reload(ld)
#
# dat = np.random.rand(30,40,50) > 0.5;
# mask = sel.structureElement('Disk', (5,5,5));
# indices = np.where(dat.reshape(-1))[0];
# c_id = len(indices)/2;
# c = indices[c_id];
# xyz = np.unravel_index(c, dat.shape)
# l = np.array(mask.shape)/2
# r = np.array(mask.shape) - l;
# dlo = [max(0,xx-ll) for xx,ll in zip(xyz,l)];
# dhi = [min(xx+rr,ss) for xx,rr,ss in zip(xyz,r, dat.shape)]
# mlo = [-min(0,xx-ll) for xx,ll in zip(xyz,l)];
# mhi = [mm + min(0, ss-xx-rr) for xx,rr,ss,mm in zip(xyz,r, dat.shape, mask.shape)]
#
# nbh = dat[dlo[0]:dhi[0], dlo[1]:dhi[1], dlo[2]:dhi[2]];
# nbhm = np.logical_and(nbh, mask[mlo[0]:mhi[0], mlo[1]:mhi[1], mlo[2]:mhi[2]] > 0);
# nxyz = np.where(nbhm);
# nxyz = [nn + dl for nn,dl in zip(nxyz, dlo)];
# nbi = np.ravel_multi_index(nxyz, dat.shape);
#
# nbs = ld.findNeighbours(indices, c_id , dat.shape, dat.strides, mask)
#
# nbs.sort();
# print np.all(nbs == nbi)
#
#
# dat = np.random.rand(30,40,50) > 0.5;
# indices = np.where(dat.reshape(-1))[0];
# c_id = len(indices)/2;
# c = indices[c_id];
# xyz = np.unravel_index(c, dat.shape)
# l = np.array([2,2,2]);
# r = l + 1;
# dlo = [max(0,xx-ll) for xx,ll in zip(xyz,l)];
# dhi = [min(xx+rr,ss) for xx,rr,ss in zip(xyz, r, dat.shape)]
# nbh = dat[dlo[0]:dhi[0], dlo[1]:dhi[1], dlo[2]:dhi[2]];
# nxyz = np.where(nbh);
# nxyz = [nn + dl for nn,dl in zip(nxyz, dlo)];
# nbi = np.ravel_multi_index(nxyz, dat.shape);
#
# nbs = ld.findNeighbours(indices, c_id , dat.shape, dat.strides, tuple(l))
#
# nbs.sort();
# print np.all(nbs == nbi)
#
# print nbs
# print nbi
#
#
# import numpy as np
# from ClearMap.Utils.Timer import Timer;
# import ClearMap.DataProcessing.LargeData as ld
# reload(ld)
#
# data = np.random.rand(100);
# values =np.random.rand(50);
# indices = np.arange(50);
# ld.setArray(data, indices, values, cutoff = 1)
# print np.all(data[:50] == values)
#
# import numpy as np
# from ClearMap.Utils.Timer import Timer;
# import ClearMap.DataProcessing.LargeData as ld
# reload(ld)
#
# m = np.array([1,3,6,7,10]);
# i = np.array([1,2,3,4,6,7,8,9]);
#
# o = ld.match(m,i)
#
# o2 = [np.where(i==l)[0][0] for l in m]
#
#
#
def _test():
#%%
import numpy as np
import scipy.ndimage as ndi
import ClearMap.DataProcessing.LargeData as ld
import ClearMap.Visualization.Plot3d as p3d
import ClearMap.DataProcessing.ConvolvePointList as cpl
import ClearMap.ImageProcessing.Skeletonization.Topology3d as t3d
import ClearMap.ImageProcessing.Skeletonization.SkeletonCleanUp as scu
import ClearMap.ImageProcessing.Tracing.Connect as con
reload(con)
data = np.load('/home/ckirst/Desktop/data.npy')
binary = np.load('/home/ckirst/Desktop/binarized.npy')
skel = np.load('/home/ckirst/Desktop/skel.npy')
#points = np.load('/home/ckirst/Desktop/pts.npy');
data = np.copy(data, order='F')
binary = np.copy(binary, order='F')
skel = np.copy(skel, order='F')
skel_copy = np.copy(skel, order='F')
points = np.ravel_multi_index(np.where(skel), skel.shape, order='F')
skel, points = scu.cleanOpenBranches(skel,
skel_copy,
points,
length=3,
clean=True)
deg = cpl.convolve3DIndex(skel, t3d.n26, points)
ends, isolated = con.findEndpoints(skel, points, border=25)
special = np.sort(np.hstack([ends, isolated]))
ends_xyz = np.array(np.unravel_index(ends, data.shape, order='F')).T
isolated_xyz = np.array(np.unravel_index(isolated, data.shape,
order='F')).T
special_xyz = np.vstack([ends_xyz, isolated_xyz])
#%%
import ClearMap.ParallelProcessing.SharedMemoryManager as smm
data_s = smm.asShared(data, order='F')
binary_s = smm.asShared(binary.view('uint8'), order='F')
skel_s = smm.asShared(skel.view('uint8'), order='F')
smm.clean()
res = con.addConnections(data_s,
binary_s,
skel_s,
points,
radius=20,
start_points=None,
add_to_skeleton=True,
add_to_mask=True,
verbose=True,
processes=4,
debug=False,
block_size=10)
skel_s = skel_s.view(bool)
binary_s = binary_s.view(bool)
#%%
mask_img = np.asarray(binary, dtype=int, order='A')
mask_img[:] = mask_img + binary_s
mask_img[:] = mask_img + skel
data_img = np.copy(data, order='A')
data_img[skel] = 120
mask_img_f = np.reshape(mask_img, -1, order='A')
data_img_f = np.reshape(data_img, -1, order='A')
mask_img_f[res] = 7
data_img_f[res] = 512
mask_img_f[special] = 8
data_img_f[special] = 150
for d in [3, 4, 5]:
mask_img_f[points[deg == d]] = d + 1
try:
con.viewer[0].setSource(mask_img)
con.viewer[1].setSource(data_img)
except:
con.viewer = p3d.plot([mask_img, data_img])
con.viewer[0].setMinMax([0, 8])
con.viewer[1].setMinMax([24, 160])
#%%
mask = binary
data_new = np.copy(data, order='A')
data_new[skel] = 120
skel_new = np.asarray(skel, dtype=int, order='A')
skel_new[:] = skel_new + binary
binary_new = np.copy(binary, order='A')
qs = []
for i, e in enumerate(special):
print('------')
print('%d / %d' % (i, len(special)))
path, quality = con.connectPoint(data,
mask,
special,
i,
radius=25,
skeleton=skel,
tubeness=None,
remove_local_mask=True,
min_quality=15.0,
verbose=True,
maxSteps=15000,
costPerDistance=1.0)
#print path, quality
if len(path) > 0:
qs.append(quality * 1.0 / len(path))
q = con.addPathToMask(skel_new, path, value=7)
q = con.addPathToMask(data_new, path, value=512)
binary_new = con.addDilatedPathToMask(binary_new,
path,
iterations=1)
skel_new[:] = skel_new + binary_new
q = con.addPathToMask(skel_new, special_xyz, value=6)
for d in [3, 4, 5]:
xyz = np.array(
np.unravel_index(points[deg == d], data.shape, order='F')).T
q = con.addPathToMask(skel_new, xyz, value=d)
q = con.addPathToMask(data_new, special_xyz, value=150)
try:
con.viewer[0].setSource(skel_new)
con.viewer[1].setSource(data_new)
except:
con.viewer = p3d.plot([skel_new, data_new])
con.viewer[0].setMinMax([0, 8])
con.viewer[1].setMinMax([24, 160])
#%%
import matplotlib.pyplot as plt
plt.figure(1)
plt.clf()
#plt.plot(qs);
plt.hist(qs)
#%%
i = 20
i = 21
i = 30
i = 40
r = 25
center = np.unravel_index(ends[i], data.shape)
print(center, data.shape)
mask = binary
path = con.tracePointToMask(data,
mask,
center,
radius=r,
points=special_xyz,
plot=True,
skel=skel,
binary=binary,
tubeness=None,
removeLocalMask=True,
maxSteps=None,
verbose=False,
costPerDistance=0.0)
#%%
nbs = ap.findNeighbours(ends, i, skel.shape, skel.strides, r)
center = np.unravel_index(ends[i], skel.shape)
nbs_xyz = np.array(np.unravel_index(nbs, skel.shape)).T
dists = nbs_xyz - center
dists = np.sum(dists * dists, axis=1)
nb = np.argmin(dists)
center = np.unravel_index(ends[i], data.shape)
print(center, data.shape)
mask = binary
path = con.tracePointToNeighbor(data,
mask,
center,
nbs_xyz[nb],
radius=r,
points=special_xyz,
plot=True,
skel=skel,
binary=binary,
tubeness=None,
removeLocalMask=True,
maxSteps=None,
verbose=False,
costPerDistance=0.0)
#%%
import ClearMap.ImageProcessing.Filter.FilterKernel as fkr
dog = fkr.filterKernel('DoG', size=(13, 13, 13))
dv.plot(dog)
data_filter = ndi.correlate(np.asarray(data, dtype=float), dog)
data_filter -= data_filter.min()
data_filter = data_filter / 3.0
#dv.dualPlot(data, data_filter);
#%%add all paths
reload(con)
r = 25
mask = binary
data_new = data.copy()
data_new[skel] = 120
skel_new = np.asarray(skel, dtype=int)
skel_new = skel_new + binary
binary_new = binary.copy()
for i, e in enumerate(special):
center = np.unravel_index(e, data.shape)
print(i, e, center)
path = con.tracePointToMask(data,
mask,
center,
radius=r,
points=special_xyz,
plot=False,
skel=skel,
binary=binary,
tubeness=None,
removeLocalMask=True,
maxSteps=15000,
costPerDistance=1.0)
q = con.addPathToMask(skel_new, path, value=7)
q = con.addPathToMask(data_new, path, value=512)
binary_new = con.addDilatedPathToMask(binary_new, path, iterations=1)
q = con.addPathToMask(skel_new, special_xyz, value=6)
for d in [3, 4, 5]:
xyz = np.array(np.unravel_index(points[deg == d], data.shape)).T
q = con.addPathToMask(skel_new, xyz, value=d)
q = con.addPathToMask(data_new, special_xyz, value=150)
skel_new = skel_new + binary_new
try:
con.viewer[0].setSource(skel_new)
con.viewer[1].setSource(data_new)
except:
con.viewer = dv.dualPlot(skel_new, data_new)
con.viewer[0].setMinMax([0, 8])
con.viewer[1].setMinMax([24, 160])
#%%
import ClearMap.ImageProcessing.Skeletonization.Skeletonize as skl
skel_2 = skl.skeletonize3D(binary_new.copy())
#%%
np.save('/home/ckirst/Desktop/binarized_con.npy', binary_new)
#%%
# write image
import ClearMap.IO.IO as io
#r = np.asarray(128 * binary_new, dtype = 'uint8');
#g = r.copy(); b = r.copy();
#r[:] = r + 127 * skel_2[0];
#g[:] = g - 128 * skel_2[0];
#b[:] = b - 128 * skel_2[0];
#img = np.stack((r,g,b), axis = 3)
img = np.asarray(128 * binary_new, dtype='uint8')
img[:] = img + 127 * skel_2[0]
io.writeData('/home/ckirst/Desktop/3d.tif', img)
def graph_from_skeleton(skeleton, points = None, radii = None, vertex_coordinates = True,
check_border = True, delete_border = False, verbose = False):
"""Converts a binary skeleton image to a graph-tool graph.
Arguments
---------
skeleton : array
Source with 2d/3d binary skeleton.
points : array
List of skeleton points as 1d indices of flat skeleton array (optional to save processing time).
radii : array
List of radii associated with each vertex.
vertex_coordinates : bool
If True, store coordiantes of the vertices / edges.
check_border : bool
If True, check if the boder is empty. The algorithm reuqires this.
delete_border : bool
If True, delete the border.
verbose : bool
If True, print progress information.
Returns
-------
graph : Graph class
The graph corresponding to the skeleton.
Note
----
Edges are detected between neighbouring foreground pixels using 26-connectivty.
"""
skeleton = io.as_source(skeleton);
if delete_border:
skeleton = t3d.delete_border(skeleton);
check_border = False;
if check_border:
if not t3d.check_border(skeleton):
raise ValueError('The skeleton array needs to have no points on the border!');
if verbose:
timer = tmr.Timer();
timer_all = tmr.Timer();
print('Graph from skeleton calculation initialize.!');
if points is None:
points = ap.where(skeleton.reshape(-1, order = 'A')).array;
if verbose:
timer.print_elapsed_time('Point list generation');
timer.reset();
#create graph
n_vertices = points.shape[0];
g = ggt.Graph(n_vertices=n_vertices, directed=False);
g.shape = skeleton.shape;
if verbose:
timer.print_elapsed_time('Graph initialized with %d vertices' % n_vertices);
timer.reset();
#detect edges
edges_all = np.zeros((0,2), dtype = int);
for i,o in enumerate(t3d.orientations()):
# calculate off set
offset = np.sum((np.hstack(np.where(o))-[1,1,1]) * skeleton.strides)
edges = ap.neighbours(points, offset);
if len(edges) > 0:
edges_all = np.vstack([edges_all, edges]);
if verbose:
timer.print_elapsed_time('%d edges with orientation %d/13 found' % (edges.shape[0], i+1));
timer.reset();
if edges_all.shape[0] > 0:
g.add_edge(edges_all);
if verbose:
timer.print_elapsed_time('Added %d edges to graph' % (edges_all.shape[0]));
timer.reset();
if vertex_coordinates:
vertex_coordinates = np.array(np.unravel_index(points, skeleton.shape, order=skeleton.order)).T;
g.set_vertex_coordinates(vertex_coordinates);
if radii is not None:
g.set_vertex_radius(radii);
if verbose:
timer_all.print_elapsed_time('Skeleton to Graph');
return g;
def temp():
import scipy.ndimage as ndi
corr_ndi = ndi.correlate1d(data,
kernel,
axis=axis,
mode='constant',
cval=0)
assert np.allclose(corr.array, corr_ndi)
c = corr.array
c[0, :, 0]
corr_ndi[0, :, 0]
data = np.array(np.random.rand(1000, 1000, 500), order='F')
data = np.array(np.random.rand(300, 400, 1500), order='F')
kernel = np.array([1, 2, 3, 4, 5])
import ClearMap.Utils.Timer as tmr
timer = tmr.Timer()
for axis in range(3):
corr = ap.correlate1d(data,
kernel,
axis=axis,
verbose=False,
processes=None)
timer.print_elapsed_time('ap')
import ClearMap.Utils.Timer as tmr
timer = tmr.Timer()
for axis in range(3):
corr2 = ap2.correlate1d(data,
kernel,
axis=axis,
verbose=False,
processes=None)
timer.print_elapsed_time('ap')
import scipy.ndimage as ndi
timer = tmr.Timer()
for axis in range(3):
corr_ndi = ndi.correlate1d(data,
kernel,
axis=axis,
mode='constant',
cval=0)
timer.print_elapsed_time('ndi')
assert np.allclose(corr.array, corr_ndi)
assert np.allclose(corr2.array, corr_ndi)
# IO
import ClearMap.ParallelProcessing.DataProcessing.ArrayProcessing as ap
import numpy as np
reload(ap)
data = np.random.rand(10, 200, 10)
sink = ap.write('test.npy', data, verbose=True)
assert (np.all(sink.array == data))
read = ap.read('test.npy', verbose=True)
assert (np.all(read.array == data))
ap.io.delete_file('test.npy')
# where
reload(ap)
data = np.random.rand(30, 20, 40) > 0.5
where_np = np.array(np.where(data)).T
where = ap.where(data, cutoff=2**0)
check_np = np.zeros(data.shape, dtype=bool)
check = np.zeros(data.shape, dtype=bool)
check_np[tuple(where_np.T)] = True
check[tuple(where.array.T)] = True
assert (np.all(check_np == check))
def skeletonize(source,
sink=None,
points=None,
method='PK12i',
steps=None,
in_place=False,
verbose=True,
**kwargs):
"""Skeletonize 3d binary arrays.
Arguments
---------
source : array or source
Binary image to skeletonize.
sink : sink specification
Optional sink.
points : array or None
Optional point list of the foreground points in the binary.
method : str
'PK12' or faster index version 'PK12i'.
steps : int or None
Number of maximal iteration steps. If None, maximal thinning.
in_place : bool
If True, the skeletonization is done directly on the input array.
Returns
-------
skeleton : Source
The skeletonized array.
"""
if verbose:
timer = tmr.Timer()
if not in_place and io.is_file(source):
binary_buffer = ap.read(source).as_buffer()
else:
binary, binary_buffer = ap.initialize_source(source)
if not in_place:
binary_buffer = np.array(binary_buffer)
if method == 'PK12':
result = PK12.skeletonize(binary_buffer,
points=points,
steps=steps,
verbose=verbose,
**kwargs)
elif method == 'PK12i':
result = PK12.skeletonize_index(binary_buffer,
points=points,
steps=steps,
verbose=verbose,
**kwargs)
else:
raise RuntimeError('Skeletonizaton method %r is not valid!' % method)
if verbose:
timer.print_elapsed_time(head='Skeletonization')
if sink is None:
sink = ap.io.as_source(result)
elif isinstance(sink, str):
sink = ap.write(sink, result)
else:
sink = io.write(sink, result)
return sink
def _test():
import numpy as np
import ClearMap.ImageProcessing.Topology.Topology3d as top
from importlib import reload
reload(top)
label = top.cube_labeled()
top.print_cube(label)
# Test rotations
c = np.zeros((3, 3, 3), dtype=bool)
c[1, 0, 0] = True
top.print_cube(c)
cs = [top.rotate(c, axis=2, steps=r) for r in range(4)]
[top.print_cube(cc) for cc in cs]
reload(top)
l = top.cube_labeled()
rts = top.rotations6(l)
[top.print_cube(r) for r in rts]
reload(top)
b = top.cube_from_index(6)
i = top.cube_to_index(b)
print(i, 6)
us = np.zeros((3, 3, 3), dtype=int)
us[1, 1, 2] = 1
us[1, 0, 1] = 1
us[1, 2, 0] = 2
r12 = top.rotations12(us)
[top.print_cube(cc) for cc in r12]
#check configuration utlity
reload(top)
index = 11607
source = top.cube_from_index(index)
c = top.index_from_binary(source)
c[1, 1, 1] == index
x = np.random.rand(1500, 500, 500) > 0.6
c = top.index_from_binary(x)
import numpy as np
import ClearMap.ImageProcessing.Topology.Topology3d as top
#check fortran vs c order
x = np.random.rand(5, 5, 5) > 0.35
y = np.asanyarray(x, order='F')
ix = top.index_from_binary(x)
iy = top.index_from_binary(y)
ax = ix.array
ay = iy.array
#%% profile
import io
io.DEFAULT_BUFFER_SIZE = 2**32
import pstats, cProfile
import numpy as np
import ClearMap.ImageProcessing.Topology.Topology3d as top
x = np.ones((3000, 500, 1000), dtype=bool, order='F')
import ClearMap.IO.IO as io
import ClearMap.ParallelProcessing.DataProcessing.ArrayProcessing as ap
ap.write('test.npy', x)
y = io.as_source('test.npy')
z = io.create('resuly.npy', shape=y.shape, order='C', dtype='uint32')
cProfile.runctx(
"c =top.index_from_binary(y, method='!shared', sink=z, verbose=True, processes=None)",
globals(), locals(), "Profile.prof")
s = pstats.Stats("Profile.prof")
s.strip_dirs().sort_stats("time").print_stats()
import mmap
mmap.ACCESS_COPY
def skeletonize_index(binary,
points=None,
steps=None,
removals=False,
radii=False,
return_points=False,
check_border=True,
delete_border=False,
verbose=True):
"""Skeletonize a binary 3d array using PK12 algorithm via index coordinates.
Arguments
---------
binary : array
Binary image to be skeletonized.
steps : int or None
Number of maximal iteration steps. If None, use maximal reduction.
removals :bool
If True, returns the steps in which the pixels in the input data
were removed.
radii :bool
If True, the estimate of the local radius is returned.
verbose :bool
If True, print progress info.
Returns
-------
skeleton : array
The skeleton of the binary input.
points : nxd array
The point coordinates of the skeleton.
"""
if verbose:
print('#############################################################')
print('Skeletonization PK12 [convolution, index]')
timer = tmr.Timer()
#TODO: make this work for any memmapable source
if not isinstance(binary, np.ndarray):
raise ValueError('Numpy array required for binary in skeletonization!')
if binary.ndim != 3:
raise ValueError('The binary array dimension is %d, 3 is required!' %
binary.ndim)
if delete_border:
binary = t3d.delete_border(binary)
check_border = False
if check_border:
if not t3d.check_border(binary):
raise ValueError(
'The binary array needs to have not points on the border!')
binary_flat = binary.reshape(-1, order='A')
# detect points
if points is None:
points = ap.where(binary_flat).array
npoints = points.shape[0]
if verbose:
timer.print_elapsed_time('Foreground points: %d' % (points.shape[0], ))
if removals is True or radii is True:
#birth = np.zeros(binary.shape, dtype = 'uint16');
order = 'C'
if binary.flags.f_contiguous:
order = 'F'
death = np.zeros(binary.shape, dtype='uint16', order=order)
deathflat = death.reshape(-1, order='A')
with_info = True
else:
with_info = False
# iterate
if steps is None:
steps = -1
step = 1
nnonrem = 0
while True:
if verbose:
print(
'#############################################################'
)
print('Iteration %d' % step)
timer_iter = tmr.Timer()
print(type(points), points.dtype, binary.dtype)
border = cpl.convolve_3d_indices_if_smaller_than(
binary, t3d.n6, points, 6)
borderpoints = points[border]
#borderids = np.nonzero(border)[0];
borderids = ap.where(border).array
keep = np.ones(len(border), dtype=bool)
if verbose:
timer_iter.print_elapsed_time('Border points: %d' %
(len(borderpoints), ))
#if info is not None:
# b = birth[borderpoints[:,0], borderpoints[:,1], borderpoints[:,2]];
# bids = b == 0;
# birth[borderpoints[bids,0], borderpoints[bids,1], borderpoints[bids,2]] = step;
# sub iterations
remiter = 0
for i in range(12):
if verbose:
print(
'-------------------------------------------------------------'
)
print('Sub-Iteration %d' % i)
timer_sub_iter = tmr.Timer()
remborder = delete[cpl.convolve_3d_indices(binary, rotations[i],
borderpoints)]
rempoints = borderpoints[remborder]
if verbose:
timer_sub_iter.print_elapsed_time('Matched points : %d' %
(len(rempoints), ))
binary_flat[rempoints] = 0
keep[borderids[remborder]] = False
rem = len(rempoints)
remiter += rem
#death times
if with_info is True:
#remo = np.logical_not(keep);
deathflat[rempoints] = 12 * step + i
if verbose:
timer_sub_iter.print_elapsed_time('Sub-Iteration %d' % (i, ))
if verbose:
print(
'-------------------------------------------------------------'
)
#update foregroud
points = points[keep]
if step % 3 == 0:
npts = len(points)
points = points[consider[cpl.convolve_3d_indices(
binary, base, points)]]
nnonrem += npts - len(points)
if verbose:
print('Non-removable points: %d' % (npts - len(points)))
if verbose:
print('Foreground points : %d' % points.shape[0])
if verbose:
print(
'-------------------------------------------------------------'
)
timer_iter.print_elapsed_time('Iteration %d' % (step, ))
step += 1
if steps >= 0 and step >= steps:
break
if remiter == 0:
break
if verbose:
print('#############################################################')
timer.print_elapsed_time('Skeletonization done')
print('Total removed: %d' % (npoints - (len(points) + nnonrem)))
print('Total remaining: %d' % (len(points) + nnonrem))
if radii is True or return_points is True:
points = ap.where(binary_flat).array
if radii is True:
#calculate average diameter as death average death of neighbourhood
radii = cpl.convolve_3d_indices(death,
t3d.n18,
points,
out_dtype='uint16')
else:
radii = None
result = [binary]
if return_points:
result.append(points)
if removals is True:
result.append(death)
if radii is not None:
result.append(radii)
if len(result) > 1:
return tuple(result)
else:
return result[0]
def skeletonize(binary,
points=None,
steps=None,
removals=False,
radii=False,
check_border=True,
delete_border=False,
return_points=False,
verbose=True):
"""Skeletonize a binary 3d array using PK12 algorithm.
Arguments
---------
binary : array
Binary image to skeletonize.
points : array or None.
Optional list of points in the binary to speed up processing.
steps : int or None
Number of maximal iteration steps (if None maximal reduction).
removals : bool
If True, returns also the steps at which the pixels in the input data
where removed.
radii : bool
If True, the estimate of the local radius is returned.
check_border : bool
If True, check if the boder is empty. The algorithm reuqires this.
delete_border : bool
If True, delete the border.
verbose : bool
If True print progress info.
Returns
-------
skeleton : array
The skeleton of the binary.
points : array
The point coordinates of the skeleton nx3
Note
----
The skeletonization is done in place on the binary. Copy the binary if
needed for further processing.
"""
if verbose:
print('#############################################################')
print('Skeletonization PK12 [convolution]')
timer = tmr.Timer()
#TODO: make this work for any memmapable source !
if not isinstance(binary, np.ndarray):
raise ValueError('Numpy array required for binary in skeletonization!')
if binary.ndim != 3:
raise ValueError('The binary array dimension is %d, 3 is required!' %
binary.ndim)
if delete_border:
binary = t3d.delete_border(binary)
check_border = False
if check_border:
if not t3d.check_border(binary):
raise ValueError(
'The binary array needs to have no points on the border!')
# detect points
#points = np.array(np.nonzero(binary)).T;
if points is None:
points = ap.where(binary).array
if verbose:
timer.print_elapsed_time(head='Foreground points: %d' %
(points.shape[0], ))
if removals is True or radii is True:
#birth = np.zeros(binary.shape, dtype = 'uint16');
death = np.zeros(binary.shape, dtype='uint16')
with_info = True
else:
with_info = False
# iterate
if steps is None:
steps = -1
step = 1
removed = 0
while True:
if verbose:
print(
'#############################################################'
)
print('Iteration %d' % step)
timer_iter = tmr.Timer()
border = cpl.convolve_3d_points(binary, t3d.n6, points) < 6
borderpoints = points[border]
borderids = np.nonzero(border)[0]
keep = np.ones(len(border), dtype=bool)
if verbose:
timer_iter.print_elapsed_time('Border points: %d' %
(len(borderpoints), ))
#if info is not None:
# b = birth[borderpoints[:,0], borderpoints[:,1], borderpoints[:,2]];
# bids = b == 0;
# birth[borderpoints[bids,0], borderpoints[bids,1], borderpoints[bids,2]] = step;
# sub iterations
remiter = 0
for i in range(12):
if verbose:
print(
'-------------------------------------------------------------'
)
print('Sub-Iteration %d' % i)
timer_sub_iter = tmr.Timer()
remborder = delete[cpl.convolve_3d_points(binary, rotations[i],
borderpoints)]
rempoints = borderpoints[remborder]
if verbose:
timer_sub_iter.print_elapsed_time('Matched points: %d' %
(len(rempoints), ))
binary[rempoints[:, 0], rempoints[:, 1], rempoints[:, 2]] = 0
keep[borderids[remborder]] = False
rem = len(rempoints)
remiter += rem
removed += rem
if verbose:
print('Deleted points: %d' % (rem))
timer_sub_iter.print_elapsed_time('Sub-Iteration %d' % (i))
#death times
if with_info is True:
#remo = np.logical_not(keep);
death[rempoints[:, 0], rempoints[:, 1],
rempoints[:, 2]] = 12 * step + i
#update foreground
points = points[keep]
if verbose:
print('Foreground points: %d' % points.shape[0])
if verbose:
print(
'-------------------------------------------------------------'
)
timer_iter.print_elapsed_time('Iteration %d' % (step, ))
step += 1
if steps >= 0 and step >= steps:
break
if remiter == 0:
break
if verbose:
print('#############################################################')
print('Total removed: %d' % (removed))
print('Total remaining: %d' % (len(points)))
timer.print_elapsed_time('Skeletonization')
result = [binary]
if return_points:
result.append(points)
if removals is True:
result.append(death)
if radii is True:
#calculate average diameter as average death of neighbourhood
radii = cpl.convolve_3d(death, np.array(t3d.n18, dtype='uint16'),
points)
result.append(radii)
if len(result) > 1:
return tuple(result)
else:
return result[0]
def fill_vessels(source, sink,
resample = None, threshold = 0.5,
network = None, dtype = 'float16', cuda = None,
processing_parameter = None,
verbose = False):
"""Fill hollow tubes via a neural network.
Arguments
---------
source : str or Source
The binary data source to fill hollow tubes in.
sink : str or Source.
The binary sink to write data to. sink is created if it does not exists.
resample : int or None
If int, downsample the data by this factor, apply network and upsample.
threshold : float or None
Apply a threshold to the result of the cnn. If None, the probability of
being foreground is returned.
network : str, Model or None
The network speicifcation. If None, the default trained network is used.
dtype : str
The dtype to use for the network. See
:func:`ClearMap.ImageProcessing.MachineLearning.Torch.to` for details.
cuda : bool or None
If True, use gpu processing. If None, automatically detect gpu.
processing_parameter : dict or None
Parameter to use for block processing.
verbose : bool
If True, print progress.
Returns
-------
network : Model
The neural network model.
"""
if verbose:
timer = tmr.Timer();
#cuda
if cuda is None:
cuda = torch.cuda.is_available();
#initialize network
network = vessel_filling_network(network=network, dtype=dtype, cuda=cuda);
if not cuda: #some functions only work as float on CPU
network = network.float();
if verbose:
timer.print_elapsed_time('Vessel filling: neural network initialized')
print(network);
print('Vessel filling: using %s' % (('gpu' if cuda else 'cpu'),))
#initialize source
source = io.as_source(source);
if verbose:
timer.print_elapsed_time('Vessel filling: source loaded');
#initialize sink
if threshold:
sink_dtype = bool;
else:
sink_dtype = dtype;
sink, sink_shape = ap.initialize_sink(sink=sink, shape=source.shape, dtype=sink_dtype, order=source.order, return_buffer=False, return_shape=True);
#resampling
if resample is not None:
maxpool = torch.nn.MaxPool3d(kernel_size=resample)
upsample = torch.nn.Upsample(mode="trilinear", scale_factor=resample, align_corners=False);
if cuda:
maxpool = maxpool.cuda();
upsample = upsample.cuda();
if dtype is not None:
maxpool = tor.to(maxpool, dtype);
upsample = tor.to(upsample, dtype);
else:
maxpool = maxpool.float();
upsample = upsample.float();
#processing
if processing_parameter is None:
processing_parameter = default_fill_vessels_processing_parameter
if processing_parameter:
processing_parameter = processing_parameter.copy();
processing_parameter.update(optimization=False);
if 'size_max' not in processing_parameter or processing_parameter['size_max'] is None:
processing_parameter['size_max'] = np.max(source.shape);
if 'size_min' not in processing_parameter:
processing_parameter['size_min'] = None;
blocks = bp.split_into_blocks(source, **processing_parameter);
else:
blocks = [source];
#process blocks
for block in blocks:
if verbose:
timer_block = tmr.Timer();
print('Vessel filling: processing block %s' % (block.info()));
#load data
data = np.array(block.array);
if data.dtype == bool:
data = data.astype('uint8');
data = torch.unsqueeze(torch.from_numpy(data), 0)
if cuda:
data = tor.to(data, dtype=dtype);
data = data.cuda();
else:
data = data.float();
if verbose:
print('Vessel filling: loaded data: %r' % (tuple(data.shape),));
#down sampleprocessing_parameter
if resample:
result = maxpool(data);
else:
result = data;
result = torch.unsqueeze(result, 1);
if verbose:
print('Vessel filling: resampled data: %r' % (tuple(result.shape),));
#fill
result = network(result);
if verbose:
print('Vessel filling: network %r' % (tuple(result.shape),));
#upsample
if resample:
result = upsample(result)
if verbose:
print('Vessel filling: upsampled %r' % (tuple(result.shape),));
#write data
sink_slicing = block.slicing;
result_shape = result.shape;
result_slicing = tuple(slice(None,min(ss, rs)) for ss,rs in zip(sink_shape, result_shape[2:]));
data_slicing = (0,) + tuple(slice(None, s.stop) for s in result_slicing);
sink_slicing = bp.blk.slc.sliced_slicing(result_slicing, sink_slicing, sink_shape);
result_slicing = (0,0) + result_slicing;
#print('result', result.shape, result_slicing, 'data', data.shape, data_slicing, 'sink', sink_shape, sink_slicing)
if threshold:
sink_prev = torch.from_numpy(np.asarray(sink[sink_slicing], dtype='uint8'));
else:
sink_prev = torch.from_numpy(sink[sink_slicing]);
if cuda:
sink_prev = sink_prev.cuda();
sink_prev = tor.to(sink_prev, dtype=dtype);
else:
sink_prev = sink_prev.float();
#print('slices:', result[result_slicing].shape, data[data_slicing].shape, sink_prev.shape)
result = torch.max(torch.max(result[result_slicing], data[data_slicing]), sink_prev);
if threshold:
result = result >= threshold;
if verbose:
print('Vessel filling: thresholded %r' % (tuple(result.shape),));
if cuda:
sink[sink_slicing] = result.data.cpu();
else:
sink[sink_slicing] = result.data;
if verbose:
print('Vessel filling: result written to %r' % (sink_slicing,));
del data, result, sink_prev;
gc.collect();
if verbose:
timer_block.print_elapsed_time('Vessel filling: processing block %s' % (block.info()));
if verbose:
timer.print_elapsed_time('Vessel filling');
return sink;
def detect_cells_block(source, parameter=default_cell_detection_parameter):
"""Detect cells in a Block."""
#initialize parameter and slicings
verbose = parameter.get('verbose', False)
if verbose:
prefix = 'Block %s: ' % (source.info(), )
total_time = tmr.Timer(prefix)
base_slicing = source.valid.base_slicing
valid_slicing = source.valid.slicing
valid_lower = source.valid.lower
valid_upper = source.valid.upper
lower = source.lower
parameter_intensity = parameter.get('intensity_detection', None)
measure_to_array = dict()
if parameter_intensity:
parameter_intensity = parameter_intensity.copy()
measure = parameter_intensity.pop('measure', [])
if measure is None:
measure = []
for m in measure:
measure_to_array[m] = None
if 'source' in measure_to_array:
measure_to_array['source'] = source
# correct illumination
parameter_illumination = parameter.get('illumination_correction', None)
if parameter_illumination:
parameter_illumination = parameter_illumination.copy()
if verbose:
timer = tmr.Timer(prefix)
hdict.pprint(parameter_illumination,
head=prefix + 'Illumination correction')
save = parameter_illumination.pop('save', None)
corrected = ic.correct_illumination(source, **parameter_illumination)
if save:
save = io.as_source(save)
save[base_slicing] = corrected[valid_slicing]
if verbose:
timer.print_elapsed_time('Illumination correction')
else:
corrected = np.array(source.array)
if 'illumination' in measure_to_array:
measure_to_array['illumination'] = corrected
#background subtraction
parameter_background = parameter.get('background_correction', None)
if parameter_background:
parameter_background = parameter_background.copy()
if verbose:
timer = tmr.Timer(prefix)
hdict.pprint(parameter_background,
head=prefix + 'Background removal')
save = parameter_background.pop('save', None)
background = remove_background(corrected, **parameter_background)
if save:
save = io.as_source(save)
save[base_slicing] = corrected[valid_slicing]
if verbose:
timer.print_elapsed_time('Illumination correction')
else:
background = corrected
del corrected
if 'background' in measure_to_array:
measure_to_array['background'] = background
# equalize
parameter_equalize = parameter.get('equalization', None)
if parameter_equalize:
parameter_equalize = parameter_equalize.copy()
if verbose:
timer = tmr.Timer(prefix)
hdict.pprint(parameter_equalize, head=prefix + 'Equalization:')
save = parameter_equalize.pop('save', None)
equalized = equalize(background, mask=None, **parameter_equalize)
if save:
save = io.as_source(save)
save[base_slicing] = equalized[valid_slicing]
if verbose:
timer.print_elapsed_time('Equalization')
else:
equalized = background
del background
if 'equalized' in measure_to_array:
measure_to_array['equalized'] = equalized
#DoG filter
parameter_dog_filter = parameter.get('dog_filter', None)
if parameter_dog_filter:
parameter_dog_filter = parameter_dog_filter.copy()
if verbose:
timer = tmr.Timer(prefix)
hdict.pprint(parameter_dog_filter, head=prefix + 'DoG filter:')
save = parameter_dog_filter.pop('save', None)
dog = dog_filter(equalized, **parameter_dog_filter)
if save:
save = io.as_source(save)
save[base_slicing] = dog[valid_slicing]
if verbose:
timer.print_elapsed_time('DoG filter')
else:
dog = equalized
del equalized
if 'dog' in measure_to_array:
measure_to_array['dog'] = dog
#Maxima detection
parameter_maxima = parameter.get('maxima_detection', None)
parameter_shape = parameter.get('shape_detection', None)
if parameter_shape or parameter_intensity:
if not parameter_maxima:
print(
prefix +
'Warning: maxima detection needed for shape and intensity detection!'
)
parameter_maxima = dict()
if parameter_maxima:
parameter_maxima = parameter_maxima.copy()
if verbose:
timer = tmr.Timer(prefix)
hdict.pprint(parameter_maxima, head=prefix + 'Maxima detection:')
save = parameter_maxima.pop('save', None)
valid = parameter_maxima.pop('valid', None)
# extended maxima
maxima = md.find_maxima(source.array,
**parameter_maxima,
verbose=verbose)
if save:
save = io.as_source(save)
save[base_slicing] = maxima[valid_slicing]
#center of maxima
if parameter_maxima['h_max']:
centers = md.find_center_of_maxima(source,
maxima=maxima,
verbose=verbose)
else:
centers = ap.where(maxima).array
if verbose:
timer.print_elapsed_time('Maxima detection')
#correct for valid region
if valid:
ids = np.ones(len(centers), dtype=bool)
for c, l, u in zip(centers.T, valid_lower, valid_upper):
ids = np.logical_and(ids, np.logical_and(l <= c, c < u))
centers = centers[ids]
del ids
del dog, maxima
results = (centers, )
#cell shape detection
if parameter_shape:
parameter_shape = parameter_shape.copy()
if verbose:
timer = tmr.Timer(prefix)
hdict.pprint(parameter_shape, head=prefix + 'Shape detection:')
save = parameter_shape.pop('save', None)
# shape detection
shape = sd.detect_shape(source,
centers,
**parameter_shape,
verbose=verbose)
if save:
save = io.as_source(save)
save[base_slicing] = shape[valid_slicing]
#size detection
max_label = centers.shape[0]
sizes = sd.find_size(shape, max_label=max_label)
valid = sizes > 0
if verbose:
timer.print_elapsed_time('Shape detection')
results += (sizes, )
else:
valid = None
shape = None
#cell intensity detection
if parameter_intensity:
parameter_intensity = parameter_intensity.copy()
if verbose:
timer = tmr.Timer(prefix)
hdict.pprint(parameter_intensity,
head=prefix + 'Intensity detection:')
if not shape is None:
r = parameter_intensity.pop('shape', 3)
if isinstance(r, tuple):
r = r[0]
for m in measure:
if shape is not None:
intensity = sd.find_intensity(measure_to_array[m],
label=shape,
max_label=max_label,
**parameter_intensity)
else:
intensity = me.measure_expression(measure_to_array[m],
centers,
search_radius=r,
**parameter_intensity,
processes=1,
verbose=False)
results += (intensity, )
if verbose:
timer.print_elapsed_time('Shape detection')
if valid is not None:
results = tuple(r[valid] for r in results)
#correct coordinate offsets of blocks
results = (results[0] + lower, ) + results[1:]
#correct shapes for merging
results = tuple(r[:, None] if r.ndim == 1 else r for r in results)
if verbose:
total_time.print_elapsed_time('Cell detection')
gc.collect()
return results
def _test():
import numpy as np
import ClearMap.ParallelProcessing.DataProcessing.ArrayProcessing as ap
## Lookup table processing
#apply_lut
x = np.random.randint(0, 100, size=(20, 30))
lut = np.arange(100) + 1
y = ap.apply_lut(x, lut)
assert np.all(y == x + 1)
#apply_lut_to_index
import ClearMap.ImageProcessing.Topology.Topology3d as t3d
kernel = t3d.index_kernel(dtype=int)
import ClearMap.ImageProcessing.Binary.Smoothing as sm
lut = sm.initialize_lookup_table()
data = np.array(np.random.rand(150, 30, 40) > 0.75, order='F')
result = ap.apply_lut_to_index(data, kernel, lut, sink=None, verbose=True)
import ClearMap.Visualization.Plot3d as p3d
p3d.plot([[data, result]])
### Correlation
#correlate1d
kernel = np.array(range(11), dtype='uint32')
data = np.array(np.random.randint(0,
2**27, (300, 400, 1500),
dtype='uint32'),
order='F')
#data = np.array(np.random.rand(3,4,5), order='F');
data = np.empty((300, 400, 1500), order='F')
kernel = np.array([1, 2, 3, 4, 5], dtype='uint8')
sink = 'test.npy'
import ClearMap.Utils.Timer as tmr
import scipy.ndimage as ndi
timer = tmr.Timer()
for axis in range(3):
print(axis)
corr_ndi = ndi.correlate1d(data, axis=axis, mode='constant', cval=0)
timer.print_elapsed_time('ndi')
timer = tmr.Timer()
for axis in range(3):
print(axis)
corr = ap.correlate1d(data,
sink=sink,
kernel=kernel,
axis=axis,
verbose=False,
processes=None)
timer.print_elapsed_time('ap')
assert np.allclose(corr.array, corr_ndi)
# IO
import ClearMap.ParallelProcessing.DataProcessing.ArrayProcessing as ap
import numpy as np
reload(ap)
data = np.random.rand(10, 200, 10)
sink = ap.write('test.npy', data, verbose=True)
assert (np.all(sink.array == data))
read = ap.read('test.npy', verbose=True)
assert (np.all(read.array == data))
ap.io.delete_file('test.npy')
# where
reload(ap)
data = np.random.rand(30, 20, 40) > 0.5
where_np = np.array(np.where(data)).T
where = ap.where(data, cutoff=2**0)
check_np = np.zeros(data.shape, dtype=bool)
check = np.zeros(data.shape, dtype=bool)
check_np[tuple(where_np.T)] = True
check[tuple(where.array.T)] = True
assert (np.all(check_np == check))
def index_from_binary(source,
sink=None,
method='shared',
dtype='uint32',
processes=None,
verbose=False):
"""Calculate the local 3x3x3 configuration in a binary source.
Note
----
The configuration kernel is separable and convolution with it
is calculated via a sequence of 1d convolutions.
"""
processes, timer = ap.initialize_processing(processes=processes,
verbose=verbose,
function='index_from_binary')
#determine configuration
source, source_buffer, source_shape, source_strides, source_order = ap.initialize_source(
source,
as_1d=True,
return_shape=True,
return_strides=True,
return_order=True)
ndim = len(source_shape)
buffer_dtype = np.result_type(source_buffer.dtype, 'uint32')
delete_files = []
if source_order == 'C':
axis_range = range(ndim - 1, -1, -1)
axis_last = 0
else:
axis_range = range(ndim)
axis_last = ndim - 1
for axis in axis_range:
if axis == axis_last:
sink, sink_buffer, sink_shape, sink_strides = ap.initialize_sink(
sink=sink,
as_1d=True,
source=source,
dtype=dtype,
return_shape=True,
return_strides=True)
else:
if method == 'shared':
_, sink_buffer, sink_shape, sink_strides = ap.initialize_sink(
sink=None,
as_1d=True,
shape=source_shape,
dtype=buffer_dtype,
order=source_order,
return_shape=True,
return_strides=True)
else:
location = tempfile.mktemp() + '.npy'
_, sink_buffer, sink_shape, sink_strides = ap.initialize_sink(
sink=location,
as_1d=True,
shape=tuple(source_shape),
dtype=buffer_dtype,
order=source_order,
return_shape=True,
return_strides=True)
delete_files.append(location)
kernel = index_kernel(axis=axis, dtype=float)
#print(source_buffer.dtype, source_buffer.shape, source_shape, source_strides, axis, sink_buffer.shape, sink_buffer.dtype, sink_strides, kernel.dtype)
ap.code.correlate_1d(source_buffer, source_shape, source_strides,
sink_buffer, sink_shape, sink_strides, kernel,
axis, processes)
source_buffer = sink_buffer
for f in delete_files:
io.delete_file(f)
ap.finalize_processing(verbose=verbose,
function='index_from_binary',
timer=timer)
return sink
def read(self, *args, **kwargs): return ap.read(self.filename(*args, **kwargs));
