def interpolate_secondaries(input_folder, output_folder, factor,
target_numprojs):
sec_numprojs = len(glob.glob(f'./{input_folder}/secondary????.mha'))
img0 = itk.imread(f'./{input_folder}/secondary0000.mha')
img_origin = itk.origin(img0)
img_spacing = itk.spacing(img0)
image_ind = 0
for projnum in range(sec_numprojs - 1):
# get
image1 = f'./{input_folder}/secondary{projnum:04d}.mha'
image2 = f'./{input_folder}/secondary{projnum+1:04d}.mha'
img1_array = itk.GetArrayFromImage(itk.imread(image1))
img2_array = itk.GetArrayFromImage(itk.imread(image2))
# save the first image
itk.imwrite(itk.imread(image1),
f'./{output_folder}/secondary{image_ind:04d}.mha')
# interpolate xfactor images between those 2 images
image_interpolate_recurrence(img1_array, img2_array, image_ind,
image_ind + factor, output_folder,
img_origin, img_spacing)
image_ind = image_ind + factor
# read the last image and save it until the target_numprojs is reached
lastimage = itk.imread(
f'./{input_folder}/secondary{sec_numprojs-1:04d}.mha')
while image_ind < target_numprojs:
itk.imwrite(lastimage,
f'./{output_folder}/secondary{image_ind:04d}.mha')
image_ind = image_ind + 1
def balanced_dataset(subjects):
positive_list = []
negative_list = []
for count, (x, y) in enumerate(subjects):
x = itk.GetArrayFromImage(x)
y = itk.GetArrayFromImage(y)
x = np.array(x)
y = np.array(y)
if y.max() == 1.0:
# input = torch.from_numpy(x).type(torch.FloatTensor)
positive_list.append((x, 1.0))
else:
negative_list.append((x, 0.0))
positive_count = len(positive_list)
negative_list_1 = random.sample(negative_list, positive_count)
balanced_list = positive_list + negative_list_1
random.shuffle(balanced_list)
return balanced_list
def makeHardClassificationPatches(self, DirPath, inputPath, maskPath,
outputPath):
imgInputPath = DirPath + "/" + inputPath #"/DATA/vascu_deepV2/maskWholeImage.nii"
imgMaskPath = DirPath + "/" + maskPath
imgROIPath = DirPath + "/" + outputPath
imgInput = itk.imread(imgInputPath)
imgMask = itk.imread(imgMaskPath)
PixelType = itk.UC
Dimension = 3
ImageType = itk.Image[PixelType, Dimension]
radiusValue = 4
StructuringElementType = itk.FlatStructuringElement[Dimension]
structuringElement = StructuringElementType.Ball(radiusValue)
DilateFilterType = itk.BinaryDilateImageFilter[ImageType, ImageType,
StructuringElementType]
dilateFilter = DilateFilterType.New()
dilateFilter.SetInput(imgMask)
dilateFilter.SetKernel(structuringElement)
dilateFilter.SetForegroundValue(255)
imgMask = dilateFilter.GetOutput()
imgInput = itk.GetArrayFromImage(imgInput)
imgMask = itk.GetArrayFromImage(imgMask)
imgROI = np.copy(imgInput)
imgROI[imgMask > 0] = 0
itk.imwrite(itk.GetImageFromArray(imgROI.astype(np.uint8)), imgROIPath)
def get_intersection_volume(roilist,xvoxel=1.,yvoxel=1.):
# There is probably a clever way to compute this by constructing
# an "intersection contour" for each layer: for each contour, keep only
# points that are inside all other contours in the list. But is tough to then
# put those points in the right order.
# Instead we'll just make a grid of points and get the volume of the combined mask.
# With xvoxel and yvoxel the caller can tweak the voxel size of the mask in x and y.
# In z the voxel size is given by the incoming ROIs.
dz = min([r.dz for r in roilist])
assert(dz>0)
assert(xvoxel>0)
assert(yvoxel>0)
bb = bounding_box(bb=roilist[0].bb)
for roi in roilist[1:]:
bb.intersect(roi.bb)
if bb.empty:
# too bad
return 0.
spacing = np.array([xvoxel,yvoxel,dz],dtype=float)
bb.add_margins(2*spacing)
dimsize = np.array(np.round((bb.maxcorner-bb.mincorner)/spacing),dtype=int)
#img = sitk.Image(dimsize,sitk.sitkUInt8)
img = itk.GetImageFromArray(np.zeros(dimsize[::-2],dtype=np.uint8))
img.SetOrigin(bb.mincorner)
img.SetSpacing(spacing)
itkmask = itk.GetArrayFromImage(roilist[0].get_mask(img))
for roi in roilist[1:]:
itkmask *= itk.GetArrayFromImage(roi.get_mask(img))
return np.sum(itkmask)*np.prod(spacing)
def merge_pred(results_folder, ROI_list):
# get the list of Task folders
tasks_folders = glob.glob(f'{results_folder}/segm_results/Task*')
if len(tasks_folders) != 3:
print('Not all direction available, no merging was performed')
sys.exit()
# get the list of labels
all_patients = glob.glob(
f'{results_folder}/segm_results/{os.path.basename(tasks_folders[0])}/predicted_labels/*'
)
merged_labels = f'{results_folder}/segm_results/merged_labels'
merged_binaries = f'{results_folder}/segm_results/merged_binaries'
maybe_mkdir_p(merged_labels)
maybe_mkdir_p(merged_binaries)
# loop through the patient
for patient in all_patients:
label_name = os.path.basename(patient)
patient_name = label_name.replace('label_', '')
merged_binaries_patient = f'{merged_binaries}/{os.path.splitext(patient_name)[0]}'
maybe_mkdir_p(merged_binaries_patient)
# load labels in each direction
label_d1_img = itk.imread(
f'{tasks_folders[0]}/predicted_labels/{label_name}')
label_origin = label_d1_img.GetOrigin()
label_spacing = label_d1_img.GetSpacing()
label_direction = label_d1_img.GetDirection()
label_d1 = itk.GetArrayFromImage(label_d1_img)
label_d2 = itk.GetArrayFromImage(
itk.imread(f'{tasks_folders[1]}/predicted_labels/{label_name}'))
label_d3 = itk.GetArrayFromImage(
itk.imread(f'{tasks_folders[2]}/predicted_labels/{label_name}'))
# get the intersection of every 2 labels
intersec1 = np.where(label_d1 == label_d2, label_d1, 0)
intersec2 = np.where(label_d1 == label_d3, label_d1, 0)
intersec3 = np.where(label_d2 == label_d3, label_d2, 0)
# get where 2 labels agree
merged_label = np.zeros(label_d1.shape)
merged_label = np.where(intersec1 != 0, intersec1, merged_label)
merged_label = np.where(intersec2 != 0, intersec2, merged_label)
merged_label = np.where(intersec3 != 0, intersec3, merged_label)
merged_label_img = itk.GetImageFromArray(merged_label)
merged_label_img.SetOrigin(label_origin)
merged_label_img.SetSpacing(label_spacing)
merged_label_img.SetDirection(label_direction)
itk.imwrite(merged_label_img, f'{merged_labels}/{label_name}')
segmap_to_binaries(merged_label, merged_binaries_patient, label_origin,
label_spacing, label_direction, ROI_list)
def isIslandWithinDistance(image,
distanceImage,
label,
maxDistance
):
values = itk.GetArrayFromImage(image)
dist_values = itk.GetArrayFromImage(distanceImage)
return np.any(dist_values[values == label] < maxDistance)
def MaskOriginal(OriginalImg, FinalSegm):
OriginalImg_np = itk.GetArrayFromImage(OriginalImg)
FinalSegm_np = itk.GetArrayFromImage(FinalSegm)
OriginalImg_np[FinalSegm_np < 1] = 0.
OriginalImg_masked = itk.GetImageFromArray(OriginalImg_np)
OriginalImg_masked.SetOrigin(OriginalImg.GetOrigin())
OriginalImg_masked.SetSpacing(OriginalImg.GetSpacing())
OriginalImg_masked.SetDirection(OriginalImg.GetDirection())
return OriginalImg_masked
def LDMap(Input1, Input2, ImageType, spacing=True):
distance_1 = itk.GetArrayFromImage(
SignedMaurerDistanceMap(Input1, ImageType, spacing))
distance_2 = itk.GetArrayFromImage(
SignedMaurerDistanceMap(Input2, ImageType, spacing))
A1 = (distance_1 > 1e-5).astype(np.float32)
B1 = (distance_2 > 1e-5).astype(np.float32)
LDMap_out = np.abs(A1 - B1) * np.maximum(distance_1, distance_2)
return LDMap_out
def RefineSegmentation(islandImage,
subIslandLabels,
ROI):
# assignIdsToPixels
IslandsValues = itk.GetArrayFromImage(islandImage)
_pixelIdImage = np.zeros(IslandsValues.shape)
roi = itk.GetArrayFromImage(ROI)
_pixelIdImage[roi==0] = -1
_totalPixelsInROI = np.sum(roi!=0)
_pixelIdImage[roi!=0] = range(_totalPixelsInROI)
# SheetnessBasedDataCost_compute for initializeDataCosts prep
# 0
dataCostSink = np.zeros(IslandsValues.shape)
cond = (IslandsValues == subIslandLabels[1]) & roi!=0
dataCostSink[cond] = 1000
# 1
dataCostSource = np.zeros(IslandsValues.shape)
cond = (IslandsValues == subIslandLabels[0]) & roi!=0
dataCostSource[cond] = 1000
dataCostPixels = _pixelIdImage[roi!=0].flatten()
flat_dataCostSink = dataCostSink[roi!=0].flatten()
flat_dataCostSource = dataCostSource[roi!=0].flatten()
# initializeNeighbours prep
Xcenters, XFromCenter, XToCenter = SmoothnessCostFunction(pixelLeft = _pixelIdImage[:, :, :-1],
pixelRight = _pixelIdImage[:, :, 1:])
Ycenters, YFromCenter, YToCenter = SmoothnessCostFunction(pixelLeft = _pixelIdImage[:, :-1, :],
pixelRight = _pixelIdImage[:, 1:, :])
Zcenters, ZFromCenter, ZToCenter = SmoothnessCostFunction(pixelLeft = _pixelIdImage[:-1,:,:],
pixelRight = _pixelIdImage[1:,:,:])
CentersPixels = np.concatenate([Zcenters[0], Ycenters[0], Xcenters[0] ])
NeighborsPixels = np.concatenate([Zcenters[1], Ycenters[1], Xcenters[1] ])
_totalNeighbors = len(NeighborsPixels)
flat_smoothCostFromCenter = np.concatenate([ZFromCenter, YFromCenter, XFromCenter ])
flat_smoothCostToCenter = np.concatenate([ZToCenter, YToCenter, XToCenter ])
# Call Maxflow
uint_gcresult = GraphCutSupport.RunGraphCut(_totalPixelsInROI,
np.ascontiguousarray(dataCostPixels, dtype=np.uint32),
np.ascontiguousarray(flat_dataCostSource, dtype=np.uint32),
np.ascontiguousarray(flat_dataCostSink, dtype=np.uint32),
_totalNeighbors,
np.ascontiguousarray(CentersPixels, dtype=np.uint32),
np.ascontiguousarray(NeighborsPixels, dtype=np.uint32),
np.ascontiguousarray(flat_smoothCostFromCenter, dtype=np.uint32),
np.ascontiguousarray(flat_smoothCostToCenter, dtype=np.uint32)
)
_labelIdImage = _pixelIdImage
_labelIdImage[roi!=0] = uint_gcresult
_labelIdImage[roi==0] = 0
_labelIdImage = np.asarray(_labelIdImage, dtype=np.uint8)
gcresult = itk.GetImageFromArray(_labelIdImage)
gcresult.SetOrigin(islandImage.GetOrigin())
gcresult.SetSpacing(islandImage.GetSpacing())
gcresult.SetDirection(islandImage.GetDirection())
return gcresult
def load_image(root_dir,
image_file,
label_file,
npoints=50000,
threshold_min=1700,
threshold_max=2700):
# using "map" to repeat for each label
# join root directory with filename(s)
data_file = os.path.join(root_dir, image_file)
label_file = os.path.join(root_dir, label_file)
# load data using itk
image = itk.imread(data_file)
label = itk.imread(label_file)
# first annotated voxel "DomainFirst" (offset)
offsets = label.GetMetaDataDictionary()["DomainFirst"]
# convert strings to integer arrays
# flip array because annotation coordinates are in reversed order
offsets = np.flip(np.array(offsets.split(" "), dtype=np.int), 0)
# transform itk data to numpy arrays
volume = itk.GetArrayFromImage(image)
# normalize to [0, 1]
volume = (volume - threshold_min).astype(
np.float) / float(threshold_max - threshold_min)
label = itk.GetArrayFromImage(label)
# apply threshold + transform voxelgrid to pointcloud
points = np.argwhere((volume >= 0.0) & (volume <= 1.0))
# sample npoints out of all points
idx = np.random.choice(points.shape[0], npoints, replace=False)
sampled = points[idx, :].astype(np.int)
target = np.zeros(npoints, dtype=np.int)
xmin = offsets[0]
ymin = offsets[1]
zmin = offsets[2]
xmax = offsets[0] + label.shape[0]
ymax = offsets[1] + label.shape[1]
zmax = offsets[2] + label.shape[2]
# search corresponding points
for i in range(npoints):
s = sampled[i, :]
# only consider points inside the bounding box
if xmin <= s[0] < xmax and ymin <= s[1] < ymax and zmin <= s[2] < zmax:
x, y, z = s - np.array([xmin, ymin, zmin])
target[i] = label[x, y, z]
return sampled, target, volume, label
def split_predictions(modelresults_folder, task_id, task_name_3D, ROI_list,
resultsdir_3Dimages, resultsdir_3Dlabels):
# copy the 3D test images and their labels
taskfolder_3D = f'{nnUNet_raw_data}/Task{task_id:03d}_{task_name_3D}'
imagesTs_foldername_3D = f'{taskfolder_3D}/imagesTs/'
labelsTs_foldername_3D = f'{taskfolder_3D}/labelsTs/'
if not os.path.exists(resultsdir_3Dimages):
shutil.copytree(imagesTs_foldername_3D, resultsdir_3Dimages)
if not os.path.exists(resultsdir_3Dlabels):
shutil.copytree(labelsTs_foldername_3D, resultsdir_3Dlabels)
# split the predictions and the real labels into binaries
binaries_folder = f'{modelresults_folder}/binaries'
predictions_folder = f'{modelresults_folder}/OUTPUT_DIRECTORY_3D'
maybe_mkdir_p(binaries_folder)
reallabels_binaries_folder = f'{modelresults_folder}/real_labels_binaries'
reallabels_folder = f'{modelresults_folder}/real_labels'
maybe_mkdir_p(reallabels_binaries_folder)
# get the list of predicted labels
all_patients = sorted(glob.glob(f'{predictions_folder}/*.nii.gz'))
# loop through the patient
for patient in all_patients:
label_name = os.path.basename(patient)
binariesfolder_patient = f'{binaries_folder}/{os.path.splitext(label_name)[0]}'
maybe_mkdir_p(binariesfolder_patient)
reallabels_binariesfolder_patient = f'{reallabels_binaries_folder}/{os.path.splitext(label_name)[0]}'
maybe_mkdir_p(reallabels_binariesfolder_patient)
label_d1_img = itk.imread(patient)
label_origin = label_d1_img.GetOrigin()
label_spacing = label_d1_img.GetSpacing()
label_direction = label_d1_img.GetDirection()
#print(patient)
segmap = itk.GetArrayFromImage(label_d1_img)
merge_predictions.segmap_to_binaries(segmap, binariesfolder_patient,
label_origin, label_spacing,
label_direction, ROI_list)
label_d1_img = itk.imread(f'{reallabels_folder}/{label_name}')
label_origin = label_d1_img.GetOrigin()
label_spacing = label_d1_img.GetSpacing()
label_direction = label_d1_img.GetDirection()
#print(patient)
segmap = itk.GetArrayFromImage(label_d1_img)
merge_predictions.segmap_to_binaries(
segmap, reallabels_binariesfolder_patient, label_origin,
label_spacing, label_direction, ROI_list)
def update(self, itkFixedImage, maskImage, current_settings=None):
"""Sets new fixed image and new mask image"""
# Load Mahfouz current settings
FixCannyThr1 = current_settings['CannyThresh1']
FixCannyThr2 = current_settings['CannyThresh2']
FixCannyAperture = 2 * current_settings['Aperture'] + 1
minImgContrast = current_settings['minImgContrast']
maxImgContrast = current_settings['maxImgContrast']
# Since the HipHop pipeline deals with itk images, itkFixedImage and maskImage
# need to be converted to numpy arrays for the Mahfouz metric.
# Get COPY OF numpy array of fixed image (not just GetArrayViewFromImage!)
self.FixedImg = itk.GetArrayFromImage(itkFixedImage)
# Get Mask
self.MaskImg = itk.GetArrayFromImage(maskImage)
# rescale 16bit into 8bit with contrast stretching
img_for_canny = self._look_up_table(self.FixedImg, minImgContrast,
maxImgContrast)
# generate canny edge fixed image
FixedEdgeImg = cv2.Canny(img_for_canny,
FixCannyThr1,
FixCannyThr2,
L2gradient=False,
apertureSize=FixCannyAperture)
# Apply mask to edge image (eliminates edges outside of ROI)
if self.useMask:
FixedEdgeImg[self.MaskImg == 0.] = 0.
# blurr edge image (rescaling to uint16 is necessary)
self.FixedEdgeImg = cv2.GaussianBlur(
FixedEdgeImg * ((np.power(2, 16) - 1) / (np.power(2, 8) - 1)),
self.EdgeGaussSize, self.EdgeGaussSigma)
# Apply mask to Fixed image
if self.useMask:
self.FixedImg[self.MaskImg == 0.] = 2.**16 - 1.
if self.saveFixedImages:
# Save fixed edge images
cv2.imwrite('FixedEdgImg' + str(self.temp_counter_fix) + '.tif',
self.FixedEdgeImg)
cv2.imwrite('FixedImg' + str(self.temp_counter_fix) + '.tif',
self._look_up_table(self.FixedImg, 0, 2**16 - 1))
self.temp_counter_fix = self.temp_counter_fix + 1
def testNormalCrop(self):
ifrom = np.array([5, 6, 7])
ito = np.array([30, 40, 50])
cropped_image = _CropImageManuallyWithNumpy(self.orig_image, ifrom,
ito)
cropped_image2 = _CropAndPadImageManuallyWithNumpy(
self.orig_image, ifrom, ito, -1024)
itk_cropped_image = _CropImageWithITK(self.orig_image, ifrom, ito)
# spacing
self.assertTrue(
np.allclose(np.array(cropped_image.GetSpacing()),
self.orig_spacing))
self.assertTrue(
np.allclose(np.array(cropped_image2.GetSpacing()),
self.orig_spacing))
self.assertTrue(
np.allclose(np.array(itk_cropped_image.GetSpacing()),
self.orig_spacing))
# origin
self.assertTrue(
np.allclose(np.array(cropped_image.GetOrigin()),
self.orig_origin + self.orig_spacing * ifrom))
self.assertTrue(
np.allclose(np.array(cropped_image2.GetOrigin()),
self.orig_origin + self.orig_spacing * ifrom))
self.assertTrue(
np.allclose(np.array(itk_cropped_image.GetOrigin()),
self.orig_origin + self.orig_spacing * ifrom))
# size
new_size = np.array(cropped_image.GetLargestPossibleRegion().GetSize())
new_size2 = np.array(
cropped_image2.GetLargestPossibleRegion().GetSize())
itk_new_size = np.array(
itk_cropped_image.GetLargestPossibleRegion().GetSize())
exp_size = ito - ifrom
self.assertTrue((new_size == exp_size).all())
self.assertTrue((new_size2 == exp_size).all())
self.assertTrue((itk_new_size == exp_size).all())
# values
new_array = itk.GetArrayFromImage(cropped_image)
new_array2 = itk.GetArrayFromImage(cropped_image2)
itk_array = itk.GetArrayFromImage(itk_cropped_image)
self.assertTrue(
(new_array == self.orig_array[ifrom[2]:ito[2], ifrom[1]:ito[1],
ifrom[0]:ito[0]]).all())
self.assertTrue(
(new_array2 == self.orig_array[ifrom[2]:ito[2], ifrom[1]:ito[1],
ifrom[0]:ito[0]]).all())
self.assertTrue(
(itk_array == self.orig_array[ifrom[2]:ito[2], ifrom[1]:ito[1],
ifrom[0]:ito[0]]).all())
def FindReferenceImage(inDataList):
"""
This find the median file between all the input files
"""
x = y = z = 0
for i in range(len(inDataList)):
dim = itk.GetArrayFromImage(itk.imread(inDataList[i])).shape
if dim[0] > x:
x = dim[0]
if dim[1] > y:
y = dim[1]
if dim[2] > z:
z = dim[2]
COM = np.zeros((x, y, z))
for i in range(len(inDataList)):
tmp = itk.GetArrayFromImage(itk.imread(inDataList[i]))
COM += np.pad(tmp, (((x - tmp.shape[0]) // 2,
(x - tmp.shape[0]) - (x - tmp.shape[0]) // 2),
((y - tmp.shape[1]) // 2,
(y - tmp.shape[1]) - (y - tmp.shape[1]) // 2),
((z - tmp.shape[2]) // 2,
(z - tmp.shape[2]) - (z - tmp.shape[2]) // 2)))
COM /= len(inDataList)
dist = np.inf
idx = 0
for i in range(len(inDataList)):
tmp = itk.GetArrayFromImage(itk.imread(inDataList[i]))
tmp_dist = np.linalg.norm(COM -
np.pad(tmp, (((x - tmp.shape[0]) // 2,
(x - tmp.shape[0]) -
(x - tmp.shape[0]) // 2),
((y - tmp.shape[1]) // 2,
(y - tmp.shape[1]) -
(y - tmp.shape[1]) // 2),
((z - tmp.shape[2]) // 2,
(z - tmp.shape[2]) -
(z - tmp.shape[2]) // 2))))
if tmp_dist < dist:
idx = i
dist = tmp_dist
print(" ")
print("############# Reference File #############")
cprint(("The reference file for rigid alignment is found"), 'cyan')
cprint(("Output Median Filename : ", inDataList[idx]), 'yellow')
print("###########################################")
print(" ")
return inDataList[idx]
def itk_to_vtk_image(key, itk_image):
dims = list(itk_image.GetLargestPossibleRegion().GetSize())
extent = []
for v in dims:
extent.append(0)
extent.append(v - 1)
values = itk.GetArrayFromImage(itk_image).flatten(order='C')
return {
'vtkClass':
'vtkImageData',
'spacing':
list(itk_image.GetSpacing()),
'origin':
list(itk_image.GetOrigin()),
'extent':
extent,
'direction':
list(
itk.GetArrayFromVnlMatrix(itk_image.GetDirection().GetVnlMatrix().
as_matrix()).flatten()),
'pointData': {
'values': values,
'dataType': _itk_image_to_type(itk_image),
'numberOfComponents': itk_image.GetNumberOfComponentsPerPixel(),
},
}
def RelabelComponents(inputImage,
outputImageType = None):
# relabel = itk.RelabelComponentImageFilter[input_type, output_type].New()
# relabel.SetInput(inputImage)
# relabel.Update()
# return relabel.GetOutput()
label_field = itk.GetArrayFromImage(inputImage)
offset = 1
max_label = int(label_field.max()) # Ensure max_label is an integer
labels, labels_counts= np.unique(label_field,return_counts=True)
labels=labels[np.argsort(labels_counts)[::-1]]
labels0 = labels[labels != 0]
new_max_label = offset - 1 + len(labels0)
new_labels0 = np.arange(offset, new_max_label + 1)
output_type = label_field.dtype
required_type = np.min_scalar_type(new_max_label)
if np.dtype(required_type).itemsize > np.dtype(label_field.dtype).itemsize:
output_type = required_type
forward_map = np.zeros(max_label + 1, dtype=output_type)
forward_map[labels0] = new_labels0
inverse_map = np.zeros(new_max_label + 1, dtype=output_type)
inverse_map[offset:] = labels0
relabeled = forward_map[label_field]
result = itk.GetImageFromArray(relabeled)
result.SetOrigin(inputImage.GetOrigin())
result.SetSpacing(inputImage.GetSpacing())
result.SetDirection(inputImage.GetDirection())
if not outputImageType is None:
s,d = itk.template(inputImage)[1]
output_type = itk.Image[outputImageType,d]
result = castImage(result, OutputType=output_type)
return result
def binaryThresholding(inputImage,
lowerThreshold,
upperThreshold,
outputImageType = None,
insideValue = 1,
outsideValue = 0):
# Old version:
# s,d = itk.template(inputImage)[1]
# input_type = itk.Image[s,d]
# output_type = input_type if outputImageType is None else itk.Image[outputImageType,d]
# thresholder = itk.BinaryThresholdImageFilter[input_type, output_type].New()
# thresholder.SetInput(inputImage)
# thresholder.SetLowerThreshold( lowerThreshold )
# thresholder.SetUpperThreshold( upperThreshold )
# thresholder.SetInsideValue(insideValue)
# thresholder.SetOutsideValue(outsideValue)
# thresholder.Update()
# return thresholder.GetOutput()
values = itk.GetArrayFromImage(inputImage)
cond = (values>=lowerThreshold) & (values<=upperThreshold)
values[ cond ] = insideValue
values[ np.logical_not(cond) ] = outsideValue
result = itk.GetImageFromArray(values)
result.SetOrigin(inputImage.GetOrigin())
result.SetSpacing(inputImage.GetSpacing())
result.SetDirection(inputImage.GetDirection())
if not outputImageType is None:
s,d = itk.template(inputImage)[1]
output_type = itk.Image[outputImageType,d]
result = castImage(result, OutputType=output_type)
return result
def multiscaleSheetness(multiScaleInput,
scales,
SmoothingImageType,
roi = None,
alpha = 0.5,
beta = 0.5,
gamma = 0.5):
if not roi is None:
roi = itk.GetArrayFromImage(roi)
multiscaleSheetness = singlescaleSheetness(singleScaleInput = multiScaleInput,
scale = scales[0],
SmoothingImageType = SmoothingImageType,
roi = roi,
alpha = alpha,
beta = beta,
gamma = gamma)
if len(scales) > 1:
for scale in scales[1:]:
singleScaleSheetness = singlescaleSheetness(multiScaleInput,
scale = scale,
SmoothingImageType = SmoothingImageType,
roi = roi,
alpha = alpha,
beta = beta,
gamma = gamma)
refinement = abs(singleScaleSheetness) > abs(multiscaleSheetness)
multiscaleSheetness[refinement] = singleScaleSheetness[refinement]
multiscaleSheetness = itk.GetImageFromArray(multiscaleSheetness.astype(np.float32))
multiscaleSheetness.SetOrigin(multiScaleInput.GetOrigin())
multiscaleSheetness.SetSpacing(multiScaleInput.GetSpacing())
multiscaleSheetness.SetDirection(multiScaleInput.GetDirection())
return multiscaleSheetness
def computeSheetnessMeasure(SheetMeasInput,
roi = None,
alpha = 0.5,
beta = 0.5,
gamma = 0.5):
if isinstance(SheetMeasInput, np.ndarray):
sortedEigs = SheetMeasInput
RsImg, EigsImg, NoNullEigs = computeRs(RsInputImg = SheetMeasInput, roi=roi)
else:
sortedEigs = itk.GetArrayFromImage(SheetMeasInput)
# Sort them by abs (already done)
# l1, l2, l3 = sortedEigs[:,:,:,0], sortedEigs[:,:,:,1], sortedEigs[:,:,:,2]
# condA = np.abs(l1) > np.abs(l2)
# l1[condA], l2[condA] = l2[condA], l1[condA]
# condB = np.abs(l2) > np.abs(l3)
# l2[condB], l3[condB] = l3[condB], l2[condB]
# condC = np.abs(l1) > np.abs(l2)
# l1[condC], l2[condC] = l2[condC], l1[condC]
# sortedEigs[:,:,:,0], sortedEigs[:,:,:,1], sortedEigs[:,:,:,2] = l1, l2, l3
RsImg, EigsImg, NoNullEigs = computeRs(RsInputImg = sortedEigs, roi=roi)
SheetnessImage = np.empty(EigsImg.shape[:-1], dtype=float)
SheetnessImage[NoNullEigs] = - np.sign( sortedEigs[NoNullEigs,2] )
tmp = 1. / (beta*beta)
SheetnessImage[NoNullEigs] *= np.exp(-RsImg[NoNullEigs,0] * RsImg[NoNullEigs,0] * tmp)
tmp = 1. / (alpha*alpha)
SheetnessImage[NoNullEigs] *= np.exp(-RsImg[NoNullEigs,1] * RsImg[NoNullEigs,1] * tmp)
tmp = 1. / (gamma*gamma)
SheetnessImage[NoNullEigs] *= np.exp(-RsImg[NoNullEigs,2] * RsImg[NoNullEigs,2] * tmp)
# SheetnessImage *= EigsImg[:,:,:,2] ScaleObjectnessMeasureOff
SheetnessImage[NoNullEigs] *= ( 1 - np.exp(-RsImg[NoNullEigs,3] * RsImg[NoNullEigs,3] * 4) )
# SheetnessImage = itk.GetImageFromArray(SheetnessImage)
return SheetnessImage
def noisyImage(self, inputPath, outputName, noiseType):
for id, i in enumerate(self.noiseLevels):
img = itk.imread(inputPath)
dat = itk.GetArrayFromImage(img)
dat = dat.astype(np.float32)
# simulated CT noise poisson + gaussian noise
if (noiseType == "poisson"):
datNoisy = 0.5 * np.random.poisson(
dat, None) + 0.5 * np.random.normal(dat, i, None)
elif (noiseType == "rician"):
datNoisy = rice.rvs(dat / i, scale=i)
else:
print("error noise type not supported")
datNoisy[datNoisy < 0] = 0
datNoisy[datNoisy > 255] = 255
# writing image on disk
noisyImg = itk.GetImageFromArray(datNoisy.astype(np.uint8))
print(i)
outputPath = outputName + "_" + str(i) + ".nii"
itk.imwrite(noisyImg, outputPath)
print(outputPath)
def get_labelmap(self, tube):
if self.input_image is None:
raise Exception('No input image?????')
f = itk.SpatialObjectToImageFilter[itk.SpatialObject[3],
itk.Image[itk.UC, 3]].New()
# same origin and spacing and dir as input_image
f.SetOrigin(self.input_image.GetOrigin())
f.SetSpacing(self.input_image.GetSpacing())
f.SetDirection(self.input_image.GetDirection())
f.SetSize(self.input_image.GetLargestPossibleRegion().GetSize())
f.SetUseObjectValue(False)
f.SetOutsideValue(0)
f.SetInsideValue(1)
f.SetInput(tube)
f.Update()
mask = f.GetOutput()
voxels = itk.GetArrayFromImage(mask)
padded = np.concatenate(([0], voxels.flatten(), [0]))
run_edges = np.diff(padded)
run_starts, = np.where(run_edges > 0)
run_stops, = np.where(run_edges < 0)
run_lengths = run_stops - run_starts
return np.array(list(zip(run_starts, run_lengths)),
dtype='uint32').flatten()
def FindReferenceImage(inDataList):
"""
This find the median file between all the input files
"""
IMG = []
DIM = []
for i in range(len(inDataList)):
tmp = itk.GetArrayFromImage(itk.imread(inDataList[i]))
IMG.append(tmp)
DIM.append(tmp.shape)
ref_dim = np.max(DIM, axis=0)
for i in range(len(inDataList)):
IMG[i] = np.pad(IMG[i], ((0, ref_dim[0] - DIM[i][0]),
(0, ref_dim[1] - DIM[i][1]),
(0, ref_dim[2] - DIM[i][2])),
mode='constant',
constant_values=0)
COM = np.sum(np.asarray(IMG), axis=0) / len(inDataList)
idx = np.argmin(np.sqrt(np.sum((np.asarray(IMG) - COM)**2,
axis=(1, 2, 3))))
print(" ")
print("############# Reference File #############")
cprint(("The reference file for rigid alignment is found"), 'green')
cprint(("Output Median Filename : ", inDataList[idx]), 'yellow')
print("###########################################")
print(" ")
return inDataList[idx]
def getImages(loader_dir, image_list, down_sample):
# get all images
all_images = []
for image_path in image_list:
if down_sample:
img = downSample(image_path)
else:
image = itk.imread(image_path, itk.F)
img = itk.GetArrayFromImage(image)
all_images.append(img)
all_images = np.array(all_images)
# get mean and std
mean_path = loader_dir + 'mean_img.npy'
std_path = loader_dir + 'std_img.npy'
if not os.path.exists(mean_path) or not os.path.exists(std_path):
mean_image = np.mean(all_images)
std_image = np.std(all_images)
np.save(mean_path, mean_image)
np.save(std_path, std_image)
else:
mean_image = np.load(mean_path)
std_image = np.load(std_path)
# normlaize
norm_images = []
for image in all_images:
norm_images.append([(image - mean_image) / std_image])
return norm_images
def main():
featureFiles = glob.glob(
'/home/samuel/Projects/Ultrasound/Spectroscopy/Data/LinearProbe1/Chicken2/SpectraIteration1Features/*Spectra.npy'
)
labelFile = '/home/samuel/Projects/Ultrasound/Spectroscopy/Data/LinearProbe1/Chicken2/ManualLabels/rf_voltage_15_freq_0007500000_2017-5-31_12-50-44_ManualLabel.mha'
features = []
for ff in featureFiles:
features.append(np.load(ff))
imageType = itk.Image[itk.UC, 3]
reader = itk.ImageFileReader[imageType].New()
reader.SetFileName(labelFile)
reader.Update()
labelImage = itk.GetArrayFromImage(reader.GetOutput()).squeeze()
print(ff)
for i in range(0, 256, 10):
fig = plt.figure(1)
plt.subplot(211)
plt.imshow(labelImage)
plt.plot([i * 8 - 56, i * 8 + 56], [55, 55], color="red")
plt.subplot(212)
plt.plot(features[0][55, i, :], color="C0")
plt.plot(features[1][55, i, :], color="C1")
plt.plot(features[2][55, i, :], color="C2")
plt.plot(features[3][55, i, :], color="C3")
plt.plot(features[4][55, i, :], color="C4")
plt.plot(features[5][55, i, :], color="C5")
fig.savefig("fig-" + str(i) + ".png")
plt.show()
def get_bounding_box(image_3d_ref, depth, xy_padding=15, z_padding=8):
image_3d_ref = image_threshold_binaray_3d(image_3d_ref, 127)
PixelType = itk.ctype("unsigned char")
ImageType = itk.Image[PixelType, 3]
itkimage = itk.GetImageFromArray(image_3d_ref.astype(np.uint8))
filter = itk.BinaryShapeKeepNObjectsImageFilter[ImageType].New()
filter.SetAttribute('NumberOfPixels')
filter.SetForegroundValue(255)
filter.SetBackgroundValue(0)
filter.SetNumberOfObjects(1)
filter.SetInput(itkimage)
filter.Update()
output = filter.GetOutput()
result_label_volume = itk.GetArrayFromImage(output)
z, y, x = result_label_volume.nonzero()
z_min = max(0, min(z) - z_padding)
z_max = min(image_3d_ref.shape[0] - 1, z_min + depth)
y_min = min(y)
y_max = max(y)
x_min = min(x)
x_max = max(x)
width = x_max - x_min
height = y_max - y_min
length = max(width, height)
start_z = z_min
end_z = z_max
start_y = max(y_min - xy_padding, 0)
end_y = min(start_y + length + 2 * xy_padding, image_3d_ref.shape[1] - 1)
start_x = max(x_min - xy_padding, 0)
end_x = min(start_x + length + 2 * xy_padding, image_3d_ref.shape[2] - 1)
bbox_size = [start_z, end_z, start_y, end_y, start_x, end_x]
return bbox_size
def itk_to_vtk(itkImage):
resample_factor = 10
array = itk.GetArrayFromImage(itkImage)
#Downsample test
spacing = itkImage.GetSpacing()
spacing[2] *= resample_factor
array = array[::resample_factor, :, :]
dims = (array.shape[1], array.shape[2], array.shape[0])
downsampled_image = itk.GetImageFromArray(array)
print(downsampled_image)
vtk_array = numpy_support.numpy_to_vtk(num_array=array.ravel(),
deep=True,
array_type=vtk.VTK_FLOAT)
print(dims, spacing)
#MN
vtkImage = vtk.vtkImageData()
vtkImage.SetDimensions(dims)
vtkImage.SetSpacing(spacing)
vtkImage.GetPointData().SetScalars(vtk_array)
return vtkImage
def readNII(dcm_path):
ImageType = itk.Image[itk.F, 3]
reader = itk.ImageSeriesReader[ImageType].New()
reader.SetFileName(dcm_path)
reader.Update()
image3d = reader.GetOutput()
return itk.GetArrayFromImage(image3d)
def image_keep_max_region(label_volume_data):
label_volume = image_threshold_binaray_3d(label_volume_data, 127)
# labels = measure.label(image_3d_ref, connectivity=2)
# props = measure.regionprops(labels)
# max_area = 0
# max_area_index = 0
# for i in range(len(props)):
# area = props[i].area
# if area > max_area:
# max_area_index = i
# max_area = area
# box = props[max_area_index].bbox
# label_volume_with_max_region = np.zeros(image_3d_ref.shape, np.uint8)
# label_volume_with_max_region[box[0]:box[3], box[1]:box[4], box[2]:box[5]] = \
# image_3d_ref[box[0]:box[3], box[1]:box[4], box[2]:box[5]]
# assert box[3] - box[0] < depth + 1
# for i in range(label_volume_with_max_region.shape[0]):
# cv.imwrite('D:\\tmp\\MIS\\other\\%d.jpg' % (i), label_volume_with_max_region[i])
PixelType = itk.ctype("unsigned char")
ImageType = itk.Image[PixelType, 3]
itkimage = itk.GetImageFromArray(label_volume.astype(np.uint8))
filter = itk.BinaryShapeKeepNObjectsImageFilter[ImageType].New()
filter.SetAttribute('NumberOfPixels')
filter.SetForegroundValue(255)
filter.SetBackgroundValue(0)
filter.SetNumberOfObjects(1)
filter.SetInput(itkimage)
filter.Update()
output = filter.GetOutput()
result_label_volume = itk.GetArrayFromImage(output)
return result_label_volume
def distanceMapByFastMarcher(image,
objectLabel,
stoppingValue,
ImageType
):
FastMarchingImageFilter = itk.FastMarchingImageFilter[ImageType, ImageType]
fastMarcher = FastMarchingImageFilter.New()
fastMarcher.SetOutputSize(image.GetLargestPossibleRegion().GetSize())
fastMarcher.SetOutputOrigin(image.GetOrigin() )
fastMarcher.SetOutputSpacing(image.GetSpacing() )
fastMarcher.SetOutputDirection(image.GetDirection() )
fastMarcher.SetSpeedConstant(1.0)
if (stoppingValue > 0):
fastMarcher.SetStoppingValue(stoppingValue)
NodeType = itk.LevelSetNode.F3
FastMarchingNodeContainer = itk.VectorContainer[itk.UI, NodeType]
TrialIndexes = np.array(np.where(itk.GetArrayFromImage(image) == objectLabel)).T
seeds = FastMarchingNodeContainer.New()
seeds.Initialize()
for Idx in TrialIndexes:
node = seeds.CreateElementAt(seeds.Size())
node.SetValue(0.)
node.SetIndex(Idx[::-1].tolist())
fastMarcher.SetTrialPoints(seeds)
fastMarcher.Update()
return fastMarcher.GetOutput()
def VisualizeReferenceSpectrum(rf_files, freq_sampling):
plt.figure(1, figsize=(5, 4))
handles = []
labels = []
for rf_file in rf_files:
ComponentType = itk.ctype('float')
Dimension = 2
ImageType = itk.VectorImage[ComponentType, Dimension]
reader = itk.ImageFileReader[ImageType].New(FileName=rf_file)
reader.Update()
image = reader.GetOutput()
arr = itk.GetArrayFromImage(image)
arr /= arr[:, :, arr.shape[2] / 3 - arr.shape[2] / 5:arr.shape[2] / 2 +
arr.shape[2] / 5].max()
freq = np.linspace(freq_sampling / 2 / arr.shape[2], freq_sampling / 2,
arr.shape[2])
ax = plt.plot(freq, arr[0, 0, :].ravel(), label=rf_file)
handles.append(ax[0])
labels.append(rf_file)
plt.xlabel('Frequency [Hz]')
plt.ylabel('Power spectrum [V]')
plt.figlegend(handles, labels, 'upper right')
plt.ylim(0.0, 1.0)
dirname = os.path.dirname(rf_files[0])
plt.savefig(os.path.join(dirname, 'ReferenceSpectrum.png'), dpi=300)
plt.show()