def example_fluid():
f = box_image(
core.StructuredGrid((256, 256), spacing=[0.05, 0.05], device=device))
d = box_image(core.StructuredGrid((256, 256),
spacing=[0.05, 0.05],
device=device),
break_width=46,
val=100.0)
f = so.Gaussian.Create(channels=1,
kernel_size=5,
sigma=2,
dim=2,
device=device)(f)
grads = so.Gradient.Create(dim=2, device=device)(f) * 100
fluid_kernel = so.FluidKernel.Create(f, device=device)
# Flow the gradients
# Compute the gradients of the x
x = core.StructuredGrid.FromGrid(f,
tensor=grads[1].unsqueeze(0),
channels=1)
x_grads = so.Gradient.Create(dim=2, device=device)(grads[1])
def _create_grid(src_surface, src_excess, grid_size, grid_device):
grid_size = torch.tensor(grid_size,
device=device,
dtype=tar_surface.vertices.dtype)
extent_verts = src_surface.vertices.clone()
for surface in src_excess:
extent_verts = torch.cat([extent_verts, surface.vertices], 0)
vert_min = extent_verts.min(0).values
vert_max = extent_verts.max(0).values
# Expand beyond the min so that we contain the entire surface - 10 % should be enough
expansion = (vert_max - vert_min) * expansion_factor
vert_min -= expansion
vert_max += expansion
# the verts are in (x,y,z) and we need (z,y,x) for volumes
vert_min = vert_min.flip(0)
vert_max = vert_max.flip(0)
# Calculate the spacing
spacing = (vert_max - vert_min) / grid_size
return core.StructuredGrid(grid_size,
spacing=spacing,
origin=vert_min,
device=grid_device,
dtype=torch.float32,
requires_grad=False)
def process_mrti_data(rabbit, out_path):
rerun = False
if os.path.exists(f'{out_path}/{rabbit}_thermal_vol.nii.gz') and not rerun:
print(f'Processing MRTI for {rabbit} ... done')
return
print(f'Processing MRTI for {rabbit} ... ', end='')
data_dir = f'/hdscratch/ucair/AcuteBiomarker/Data/mrti/'
files = sorted(glob.glob(f'{data_dir}/*'))
mat_file = [x for x in files if rabbit in x][0]
mrti_dict = loadmat(mat_file)
cs_data = np.transpose(mrti_dict['PosDCS'][:, :, :, 0:3], (2, 1, 0, 3))
mrti_data = np.transpose(mrti_dict['temps'], (2, 1, 0, 3))
ctd_map = np.nan_to_num(np.transpose(mrti_dict['CTD_Map'], (2, 1, 0)),
nan=1.0)
ctd_map = np.log(ctd_map)
origin = np.ascontiguousarray(cs_data[0, 0, 0, ::-1])
thermal_vol = core.StructuredGrid(size=mrti_data.shape[0:3],
origin=origin,
spacing=torch.tensor([1.0, 1.0, 1.0]),
tensor=torch.tensor(mrti_data).permute(
3, 0, 1, 2),
channels=mrti_data.shape[-1])
log_ctd_map = core.StructuredGrid(
size=mrti_data.shape[0:3],
origin=origin,
spacing=torch.tensor([1.0, 1.0, 1.0]),
tensor=torch.tensor(ctd_map).unsqueeze(0),
channels=1)
io.SaveITKFile(thermal_vol, f'{out_path}/{rabbit}_thermal_vol.nii.gz')
io.SaveITKFile(log_ctd_map, f'{out_path}/{rabbit}_log_ctd_map.nii.gz')
print('done')
def _generate_stir_vol(stir_center, stir_space, stir_data):
# z_space = []
# for i, c in enumerate(stir_center[1:]):
# z_space.append(c[2] - stir_center[i][2])
# z_space = 2.0
# z_slab = np.round(np.array(z_space).mean(), 3)
z_slab = 2.0
z_org = np.array(stir_center)[:, 2].min()
vol_org = [
stir_center[0][0] - ((stir_data.shape[0] / 2) * stir_space[0]),
stir_center[0][1] - ((stir_data.shape[1] / 2) * stir_space[1]),
z_org + z_slab / 2.0
] # Plus or minus?
vol_space = [stir_space[0], stir_space[1], z_slab]
# stir_vol = core.StructuredGrid(
# [stir_data.shape[0], stir_data.shape[1], stir_data.shape[2]],
# spacing=vol_space,
# origin=vol_org,
# tensor=torch.tensor(stir_data).unsqueeze(0),
# channels=1
# )
stir_vol = core.StructuredGrid(
[stir_data.shape[0], stir_data.shape[2], stir_data.shape[1]],
spacing=[vol_space[0], vol_space[2], vol_space[1]],
origin=[vol_org[0], vol_org[2], vol_org[1]],
tensor=torch.tensor(stir_data.transpose(0, 2,
1)).unsqueeze(0).flip(-2),
channels=1)
return stir_vol
def exvivo_to_stacked_blocks(rabbit, block, direction, affine_only=False, base_dir='/hdscratch/ucair/'):
# This is registered from blocks to exvivo, so phi is needed to bring the exvivo MR image to the block images
# Need to determine the grid to sample the MR onto
# rabbit_dir = f'/hdscratch/ucair/{rabbit}/blockface/'
block_dir = f'{base_dir}{rabbit}/blockface/{block}/'
exvivo_dir = f'{base_dir}{rabbit}/mri/exvivo/'
# block_list = sorted(glob.glob(f'{rabbit_dir}block*'))
# orig_dir = f'/home/sci/blakez/ucair/{rabbit}/rawVolumes/ExVivo_2018-07-26/'
# Load the affine
# try:
# aff = np.loadtxt(f'{exvivo_dir}/surfaces/raw/blocks_to_exvivo_affine.txt')
# aff = torch.tensor(aff, device=device, dtype=torch.float32)
# except IOError:
aff = np.loadtxt(os.path.normpath(f'{block_dir}../recons/surfaces/raw/blocks_to_exvivo_affine.txt'))
aff = torch.tensor(aff, device=device, dtype=torch.float32)
# if affine_only:
# if direction == 'phi':
# return aff.inverse()
# else:
# return aff
# Load the deformation
# try:
# deformation = io.LoadITKFile(
# f'{exvivo_dir}/volumes/raw/blocks_{direction}_to_exvivo.mhd', device=device
# )
# deformation.set_size((256, 256, 256))
# except RuntimeError:
deformation = io.LoadITKFile(
os.path.normpath(f'{block_dir}../recons/surfaces/raw/blocks_{direction}_to_exvivo.mhd'), device=device
)
deformation.set_size((256, 256, 256))
if affine_only:
deformation.set_to_identity_lut_()
if direction == 'phi':
spacing = []
origin = []
size = []
# if 'block07' in block_path:
# hdr = tools.read_mhd_header(f'{block_path}/volumes/raw/difference_volume.mhd')
# else:
hdr = tools.read_mhd_header(f'{block_dir}/volumes/raw/{block}_phi_inv_stacking.mhd')
spacing.append(np.array([float(x) for x in hdr['ElementSpacing'].split(' ')]))
origin.append(np.array([float(x) for x in hdr['Offset'].split(' ')]))
size.append(np.array([float(x) for x in hdr['DimSize'].split(' ')]))
spacing = np.stack(spacing)
origin = np.stack(origin)
size = np.stack(size)
extent = size * spacing + origin
aff_grid_size = np.array((512, 512, 512))
aff_grid_origin = np.min(origin, axis=0)
aff_grid_spacing = (np.max(extent, axis=0) - aff_grid_origin) / aff_grid_size
aff_grid = core.StructuredGrid(
size=aff_grid_size.tolist()[::-1],
spacing=aff_grid_spacing.tolist()[::-1],
origin=aff_grid_origin.tolist()[::-1],
device=device,
channels=3
)
aff_grid.set_size(size=(512, 512, 512), inplace=True)
aff_grid.set_to_identity_lut_()
# Apply the FORWARD affine to the deformation
a = aff[0:3, 0:3].float()
t = aff[-0:3, 3].float()
# Create a deformation from the affine that lives in the stacked blocks space
aff_grid.data = aff_grid.data.flip(0)
aff_grid.data = torch.matmul(a, aff_grid.data.permute(list(range(1, 3 + 1)) + [0]).unsqueeze(-1))
aff_grid.data = (aff_grid.data.squeeze() + t).permute([-1] + list(range(0, 3)))
aff_grid.data = aff_grid.data.flip(0)
# Compose the grids
exvivo_to_blocks = so.ComposeGrids.Create(device=device)([aff_grid, deformation])
else:
deformation.data = deformation.data.flip(0)
# Apply the inverse affine to the grid
aff = aff.inverse()
a = aff[0:3, 0:3].float()
t = aff[-0:3, 3].float()
deformation.data = torch.matmul(a.unsqueeze(0).unsqueeze(0),
deformation.data.permute(list(range(1, 3 + 1)) + [0]).unsqueeze(-1))
deformation.data = (deformation.data.squeeze() + t).permute([-1] + list(range(0, 3)))
# Flip phi_inv back to the way it was
deformation.data = deformation.data.flip(0)
exvivo_to_blocks = deformation.copy()
return exvivo_to_blocks
def block_stacking(rabbit):
rabbit_dir = f'/hdscratch2/{rabbit}/blockface/'
raw_ext = '/surfaces/raw/'
rigid_ext = '/surfaces/rigid/'
deform_ext = '/surfaces/deformable/'
vol_ext = '/volumes/raw/'
stitch_ext = '/surfaces/raw/stitching/deformable/'
# Get a list of the blocks
block_list = sorted(glob.glob(f'{rabbit_dir}block*'))
# for block_path in block_list:
# block = block_path.split('/')[-1]
# with open(f'{rabbit_dir}{block}{raw_ext}{block}_deformable_config.yaml', 'r') as f:
# params = yaml.load(f, Loader=yaml.FullLoader)
# params['propagation_sigma'] = [8.0, 8.0, 3.0]
# params['currents_sigma'] = [3.0, 1.5]
# with open(f'{rabbit_dir}{block}{raw_ext}{block}_deformable_config.yaml', 'w') as f:
# yaml.dump(params, f)
# Determine the middle block
# middle_block = block_list[9]
middle_block = block_list[5]
foot_blocks = block_list[block_list.index(middle_block):]
head_blocks = block_list[:block_list.index(middle_block) + 1][::-1]
rerun = True
rigid = False
# skip_blocks = ['block01', 'block02', 'block03', 'block04', 'block05', 'block06', 'block07', 'block08']
skip_blocks = []
# Go ahead and just delete any deformable surfaces to make sure that everything resets
# for block_path in block_list:
# filelist = glob.glob(f'{block_path}/surfaces/deformable/*')
# for f in filelist:
# os.remove(f)
if rerun or not os.path.exists(
f'{middle_block}{vol_ext}{middle_block.split("/")[-1]}_phi_stacking.mhd'
):
mid_block = middle_block.split('/')[-1]
# middle = io.LoadITKFile(f'{middle_block}{vol_ext}difference_volume.mhd')
middle = io.LoadITKFile(
f'{middle_block}{vol_ext}difference_volume.nii.gz')
# middle = io.LoadITKFile(f'{middle_block}{vol_ext}{mid_block}_volume.nrrd')
middle.set_size((60, 1024, 1024))
deformation = core.StructuredGrid.FromGrid(middle, channels=3)
deformation.set_to_identity_lut_()
io.SaveITKFile(
deformation,
f'{middle_block}{vol_ext}{middle_block.split("/")[-1]}_phi_inv_stacking.mhd'
)
io.SaveITKFile(
deformation,
f'{middle_block}{vol_ext}{middle_block.split("/")[-1]}_phi_stacking.mhd'
)
affine_tform = torch.eye(4)
np.savetxt(
f'{rabbit_dir}{mid_block}{raw_ext}{mid_block}_rigid_tform.txt',
affine_tform.numpy())
np.savetxt(
f'{rabbit_dir}{mid_block}{vol_ext}{mid_block}_rigid_tform.txt',
affine_tform.numpy())
# Copy the files from raw to deformable for the middle surface
# if os.path.exists(f'{rabbit_dir}{mid_block}{stitch_ext}'):
# mid_path = f'{mid_block}{stitch_ext}'
# else:
mid_path = f'{mid_block}{raw_ext}'
files = [
f'{rabbit_dir}{mid_path}{mid_block}_decimate.obj',
f'{rabbit_dir}{mid_path}{mid_block}_ext.obj',
]
if os.path.exists(f'{rabbit_dir}{mid_path}{mid_block}_foot.obj'):
files += [f'{rabbit_dir}{mid_path}{mid_block}_foot.obj']
if os.path.exists(f'{rabbit_dir}{mid_path}{mid_block}_head.obj'):
files += [f'{rabbit_dir}{mid_path}{mid_block}_head.obj']
if os.path.exists(
f'{rabbit_dir}{mid_path}{mid_block}_foot_support.obj'):
files += [f'{rabbit_dir}{mid_path}{mid_block}_foot_support.obj']
if os.path.exists(
f'{rabbit_dir}{mid_path}{mid_block}_head_support.obj'):
files += [f'{rabbit_dir}{mid_path}{mid_block}_head_support.obj']
out_names = []
out_path = f'{rabbit_dir}{mid_block}{deform_ext}{mid_block}'
for path in files:
name = path.split('/')[-1].split(f'{mid_block}')[-1].replace(
'.', '_deformable.')
out_names += [f'{out_path}{name}']
if not os.path.exists(out_path):
os.makedirs(out_path)
for in_file, out_file in zip(files, out_names):
shutil.copy(in_file, out_file)
# Loop over the foot blocks
for i, block_path in enumerate(foot_blocks, 1):
if i == len(foot_blocks):
break
target_block = block_path.split('/')[-1]
source_block = foot_blocks[i].split('/')[-1]
if source_block in skip_blocks:
continue
# if os.path.exists(f'{rabbit_dir}{source_block}{stitch_ext}'):
# mid_path = f'{source_block}{stitch_ext}'
# else:
mid_path = f'{source_block}{raw_ext}'
target_surface_path = f'{rabbit_dir}{target_block}{deform_ext}{target_block}_foot_deformable.obj'
source_surface_path = f'{rabbit_dir}{mid_path}{source_block}_head.obj'
# if os.path.exists(f'{rabbit_dir}{mid_path}{source_block}_head_stitched.obj'):
# source_surface_path = f'{rabbit_dir}{source_block}{raw_ext}{source_block}_head_stitched.obj'
extras_paths = [
f'{rabbit_dir}{mid_path}{source_block}_decimate.obj',
f'{rabbit_dir}{mid_path}{source_block}_ext.obj',
]
if i < len(foot_blocks) - 1:
extras_paths += [f'{rabbit_dir}{mid_path}{source_block}_foot.obj']
if os.path.exists(
f'{rabbit_dir}{mid_path}{source_block}_foot_support.obj'):
extras_paths += [
f'{rabbit_dir}{mid_path}{source_block}_foot_support.obj'
]
if os.path.exists(
f'{rabbit_dir}{source_block}{deform_ext}{source_block}_head_deformable.obj'
) and not rerun:
print(
f'The deformed surface for {source_block} already exists ... Next block'
)
continue
try:
verts, faces = io.ReadOBJ(target_surface_path)
tar_surface = core.TriangleMesh(verts, faces)
tar_surface.to_(device)
except IOError:
print(
f'The deformed foot surface for {target_block} was not found ... Next block'
)
continue
# Need to see if the target needs any support
support_block = block_list[block_list.index(block_path) -
1].split('/')[-1]
if os.path.exists(
f'{rabbit_dir}{support_block}{deform_ext}{support_block}_foot_support_deformable.obj'
):
verts, faces = io.ReadOBJ(
f'{rabbit_dir}{support_block}{deform_ext}{support_block}_foot_support_deformable.obj'
)
tar_surface.add_surface_(verts.to(device=device),
faces.to(device=device))
try:
verts, faces = io.ReadOBJ(source_surface_path)
src_surface = core.TriangleMesh(verts, faces)
src_surface.to_(device)
src_surface.flip_normals_()
except IOError:
print(
f'The raw head surface for {source_block} was not found ... Next block'
)
continue
extra_surfaces = []
for path in extras_paths:
try:
verts, faces = io.ReadOBJ(path)
except IOError:
extra_name = path.split('/')[-1]
print(
f'{extra_name} not found as an extra ... removing from list'
)
_ = extras_paths.pop(extras_paths.index(path))
continue
extra_surfaces += [core.TriangleMesh(verts, faces)]
extra_surfaces[-1].to_(device)
# Load or create the dictionary for registration
try:
if rerun:
raise IOError
with open(
f'{rabbit_dir}{source_block}{raw_ext}{source_block}_affine_config.yaml',
'r') as f:
params = yaml.load(f, Loader=yaml.FullLoader)
except IOError:
params = {
'spatial_sigma': [2.0, 0.5],
'affine_lr': 1.0e-08,
'translation_lr': 1.0e-05,
'converge': 1.0,
'rigid_transform': True
}
print(f'Registering {source_block} to {target_block}:')
affine_tform = tools.affine_register(
tar_surface.copy(),
src_surface.copy(),
spatial_sigma=params['spatial_sigma'],
affine_lr=params['affine_lr'],
translation_lr=params['translation_lr'],
rigid=params['rigid_transform'],
converge=params['converge'],
device=device)
# Apply the affine to the source element and the excess
aff_tformer = uo.AffineTransformSurface.Create(affine_tform,
device=device)
aff_src_surface = aff_tformer(src_surface)
aff_extra_surface = []
for surface in extra_surfaces:
aff_extra_surface += [aff_tformer(surface)]
out_path = f'{rabbit_dir}{source_block}{rigid_ext}{source_block}'
if not os.path.exists(out_path):
os.makedirs(out_path)
for extra_path, extra_surface in zip(extras_paths, aff_extra_surface):
name = extra_path.split('/')[-1].split(
f'{source_block}')[-1].replace('.', '_rigid.')
# if '_stitched' in name:
# name = name.replace('_stitched', '')
if not os.path.exists(f'{out_path}{name}') or rerun:
io.WriteOBJ(extra_surface.vertices, extra_surface.indices,
f'{out_path}{name}')
if rigid:
continue
# Save out the parameters:
with open(
f'{rabbit_dir}{source_block}{raw_ext}{source_block}_affine_config.yaml',
'w') as f:
yaml.dump(params, f)
# Save out all of the affine transformed surfaces and the transformation
out_path = f'{rabbit_dir}{source_block}{rigid_ext}{source_block}'
# Save the affine in the volumes and in the surfaces location
np.savetxt(
f'{rabbit_dir}{source_block}{raw_ext}{source_block}_rigid_tform.txt',
affine_tform.numpy())
np.savetxt(
f'{rabbit_dir}{source_block}{vol_ext}{source_block}_rigid_tform.txt',
affine_tform.numpy())
if not os.path.exists(f'{out_path}_head_rigid.obj') or rerun:
io.WriteOBJ(aff_src_surface.vertices, aff_src_surface.indices,
f'{out_path}_head_rigid.obj')
for extra_path, extra_surface in zip(extras_paths, aff_extra_surface):
name = extra_path.split('/')[-1].split(
f'{source_block}')[-1].replace('.', '_rigid.')
if not os.path.exists(f'{out_path}{name}') or rerun:
io.WriteOBJ(extra_surface.vertices, extra_surface.indices,
f'{out_path}{name}')
try:
if rerun:
raise IOError
with open(
f'{rabbit_dir}{source_block}{raw_ext}{source_block}_deformable_config.yaml',
'r') as f:
params = yaml.load(f, Loader=yaml.FullLoader)
except IOError:
params = {
'currents_sigma': [3.0, 0.5],
'propagation_sigma': [6.0, 6.0, 3.0],
'deformable_lr': [1.0e-04, 0.5e-04],
'converge': 1.0,
'grid_size': [20, 100, 100],
'niter': 500
}
# params = {
# 'currents_sigma': [3.0, 1.5],
# 'propagation_sigma': [6.0, 6.0, 3.0],
# 'deformable_lr': [2.0e-04, 1.0e-04],
# 'converge': 1.0,
# 'grid_size': [20, 100, 100],
# 'niter': 500
# }
# Account for some old parameter names - can be deleted later
if 'spatial_sigma' in params.keys():
params['currents_sigma'] = params['spatial_sigma']
del params['spatial_sigma']
if 'phi_inv_size' in params.keys():
params['grid_size'] = params['phi_inv_size']
del params['phi_inv_size']
if 'rigid_transform' in params.keys():
del params['rigid_transform']
if 'smoothing_sigma' in params.keys():
params['propagation_sigma'] = params['smoothing_sigma']
del params['smoothing_sigma']
if type(params['deformable_lr']) is not list:
params['deformable_lr'] = [params['deformable_lr']] * len(
params['spatial_sigma'])
# Do the deformable registration
def_surface, def_extras, phi, phi_inv = tools.deformable_register(
tar_surface.copy(),
aff_src_surface.copy(),
currents_sigma=params['currents_sigma'],
prop_sigma=params['propagation_sigma'],
deformable_lr=params['deformable_lr'],
converge=params['converge'],
grid_size=params['grid_size'],
src_excess=aff_extra_surface,
accu_forward=True,
accu_inverse=True,
device=device,
grid_device='cuda:1',
iters=params['niter'])
# Save out the parameters:
with open(
f'{rabbit_dir}{source_block}{raw_ext}{source_block}_deformable_config.yaml',
'w') as f:
yaml.dump(params, f)
# Save out all of the deformable transformed surfaces and phi inv
io.SaveITKFile(
phi_inv,
f'{rabbit_dir}{source_block}{vol_ext}{source_block}_phi_inv_stacking.mhd'
)
io.SaveITKFile(
phi,
f'{rabbit_dir}{source_block}{vol_ext}{source_block}_phi_stacking.mhd'
)
out_path = f'{rabbit_dir}{source_block}{deform_ext}{source_block}'
if not os.path.exists(out_path):
os.makedirs(out_path)
if not os.path.exists(f'{out_path}_head_deformable.obj') or rerun:
io.WriteOBJ(def_surface.vertices, def_surface.indices,
f'{out_path}_head_deformable.obj')
for extra_path, extra_surface in zip(extras_paths, def_extras):
name = extra_path.split('/')[-1].split(
f'{source_block}')[-1].replace('.', '_deformable.')
# if '_stitched' in name:
# name = name.replace('_stitched', '')
if not os.path.exists(f'{out_path}{name}') or rerun:
io.WriteOBJ(extra_surface.vertices, extra_surface.indices,
f'{out_path}{name}')
print('Done registering foots blocks to middle block.')
# Loop over the head blocks
for i, block_path in enumerate(head_blocks, 1):
if i == len(head_blocks):
break
target_block = block_path.split('/')[-1]
source_block = head_blocks[i].split('/')[-1]
if source_block in skip_blocks:
continue
if os.path.exists(f'{rabbit_dir}{source_block}{stitch_ext}'):
mid_path = f'{source_block}{stitch_ext}'
else:
mid_path = f'{source_block}{raw_ext}'
target_surface_path = f'{rabbit_dir}{target_block}{deform_ext}{target_block}_head_deformable.obj'
source_surface_path = f'{rabbit_dir}{mid_path}{source_block}_foot.obj'
if os.path.exists(
f'{rabbit_dir}{mid_path}{source_block}_foot_stitched.obj'):
source_surface_path = f'{rabbit_dir}{mid_path}{source_block}_foot_stitched.obj'
extras_paths = [
f'{rabbit_dir}{mid_path}{source_block}_decimate.obj',
f'{rabbit_dir}{mid_path}{source_block}_ext.obj'
]
if i < len(head_blocks) - 1:
extras_paths += [f'{rabbit_dir}{mid_path}{source_block}_head.obj']
if os.path.exists(
f'{rabbit_dir}{mid_path}{source_block}_head_support.obj'):
extras_paths += [
f'{rabbit_dir}{mid_path}{source_block}_head_support.obj'
]
if os.path.exists(
f'{rabbit_dir}{source_block}{deform_ext}{source_block}_foot_deformable.obj'
) and not rerun:
print(
f'The deformed surface for {source_block} already exists ... Next block'
)
continue
try:
verts, faces = io.ReadOBJ(target_surface_path)
tar_surface = core.TriangleMesh(verts, faces)
tar_surface.to_(device)
except IOError:
print(
f'The deformed foot surface for {target_block} was not found ... Next block'
)
continue
# Need to see if the target needs any support
support_block = block_list[block_list.index(block_path) +
1].split('/')[-1]
if os.path.exists(
f'{rabbit_dir}{support_block}{deform_ext}{support_block}_head_support_deformable.obj'
):
verts, faces = io.ReadOBJ(
f'{rabbit_dir}{support_block}{deform_ext}{support_block}_head_support_deformable.obj'
)
tar_surface.add_surface_(verts.to(device=device),
faces.to(device=device))
try:
verts, faces = io.ReadOBJ(source_surface_path)
src_surface = core.TriangleMesh(verts, faces)
src_surface.to_(device)
src_surface.flip_normals_()
except IOError:
print(
f'The raw foot surface for {source_block} was not found ... Next block'
)
continue
extra_surfaces = []
for path in extras_paths:
try:
verts, faces = io.ReadOBJ(path)
except IOError:
extra_name = path.split('/')[-1]
print(
f'{extra_name} not found as an extra ... removing from list'
)
_ = extras_paths.pop(extras_paths.index(path))
continue
extra_surfaces += [core.TriangleMesh(verts, faces)]
extra_surfaces[-1].to_(device)
# Load or create the dictionary for registration
try:
if rerun:
raise IOError
with open(
f'{rabbit_dir}{source_block}{raw_ext}{source_block}_affine_config.yaml',
'r') as f:
params = yaml.load(f, Loader=yaml.FullLoader)
except IOError:
params = {
'spatial_sigma': [2.0, 0.5],
'affine_lr': 1.0e-08,
'translation_lr': 1.0e-05,
'converge': 1.0,
'rigid_transform': True
}
print(f'Registering {source_block} to {target_block}:')
affine_tform = tools.affine_register(
tar_surface.copy(),
src_surface.copy(),
spatial_sigma=params['spatial_sigma'],
affine_lr=params['affine_lr'],
translation_lr=params['translation_lr'],
rigid=params['rigid_transform'],
converge=params['converge'],
device=device)
# Apply the affine to the source element and the excess
aff_tformer = uo.AffineTransformSurface.Create(affine_tform,
device=device)
aff_src_surface = aff_tformer(src_surface)
aff_extra_surface = []
for surface in extra_surfaces:
aff_extra_surface += [aff_tformer(surface)]
out_path = f'{rabbit_dir}{source_block}{rigid_ext}{source_block}'
if not os.path.exists(out_path):
os.makedirs(out_path)
for extra_path, extra_surface in zip(extras_paths, aff_extra_surface):
name = extra_path.split('/')[-1].split(
f'{source_block}')[-1].replace('.', '_rigid.')
if '_stitched' in name:
name = name.replace('_stitched', '')
if not os.path.exists(f'{out_path}{name}') or rerun:
io.WriteOBJ(extra_surface.vertices, extra_surface.indices,
f'{out_path}{name}')
if rigid:
continue
# Save out the parameters:
with open(
f'{rabbit_dir}{source_block}{raw_ext}{source_block}_affine_config.yaml',
'w') as f:
yaml.dump(params, f)
# Save out all of the affine transformed surfaces and the transformation
out_path = f'{rabbit_dir}{source_block}{rigid_ext}{source_block}'
# Save the affine in the volumes and in the surfaces location
np.savetxt(
f'{rabbit_dir}{source_block}{raw_ext}{source_block}_rigid_tform.txt',
affine_tform.numpy())
np.savetxt(
f'{rabbit_dir}{source_block}{vol_ext}{source_block}_rigid_tform.txt',
affine_tform.numpy())
if not os.path.exists(f'{out_path}_foot_rigid.obj') or rerun:
io.WriteOBJ(aff_src_surface.vertices, aff_src_surface.indices,
f'{out_path}_foot_rigid.obj')
for extra_path, extra_surface in zip(extras_paths, aff_extra_surface):
name = extra_path.split('/')[-1].split(
f'{source_block}')[-1].replace('.', '_rigid.')
if not os.path.exists(f'{out_path}{name}') or rerun:
io.WriteOBJ(extra_surface.vertices, extra_surface.indices,
f'{out_path}{name}')
try:
if rerun:
raise IOError
with open(
f'{rabbit_dir}{source_block}{raw_ext}{source_block}_deformable_config.yaml',
'r') as f:
params = yaml.load(f, Loader=yaml.FullLoader)
except IOError:
params = {
'currents_sigma': [3.0, 0.5],
'propagation_sigma': [6.0, 6.0, 3.0],
'deformable_lr': [1.0e-04, 0.5e-04],
'converge': 1.0,
'grid_size': [20, 100, 100],
'niter': 500
}
# if 'spatial_sigma' in params.keys():
# params['currents_sigma'] = params['spatial_sigma']
# del params['spatial_sigma']
# if 'phi_inv_size' in params.keys():
# params['grid_size'] = params['phi_inv_size']
# del params['phi_inv_size']
# if 'rigid_transform' in params.keys():
# del params['rigid_transform']
# if 'smoothing_sigma' in params.keys():
# params['propagation_sigma'] = params['smoothing_sigma']
# del params['smoothing_sigma']
#
# if type(params['deformable_lr']) is not list:
# params['deformable_lr'] = [params['deformable_lr']] * len(params['spatial_sigma'])
# Do the deformable registration
def_surface, def_extras, phi, phi_inv = tools.deformable_register(
tar_surface.copy(),
aff_src_surface.copy(),
currents_sigma=params['currents_sigma'],
prop_sigma=params['propagation_sigma'],
deformable_lr=params['deformable_lr'],
converge=params['converge'],
grid_size=params['grid_size'],
src_excess=aff_extra_surface,
accu_forward=True,
accu_inverse=True,
device=device,
grid_device='cuda:1',
iters=params['niter'])
# Save out the parameters:
with open(
f'{rabbit_dir}{source_block}{raw_ext}{source_block}_deformable_config.yaml',
'w') as f:
yaml.dump(params, f)
# Save out all of the deformable transformed surfaces and phi inv
io.SaveITKFile(
phi_inv,
f'{rabbit_dir}{source_block}{vol_ext}{source_block}_phi_inv_stacking.mhd'
)
io.SaveITKFile(
phi,
f'{rabbit_dir}{source_block}{vol_ext}{source_block}_phi_stacking.mhd'
)
out_path = f'{rabbit_dir}{source_block}{deform_ext}{source_block}'
if not os.path.exists(out_path):
os.makedirs(out_path)
if not os.path.exists(f'{out_path}_foot_deformable.obj') or rerun:
io.WriteOBJ(def_surface.vertices, def_surface.indices,
f'{out_path}_foot_deformable.obj')
for extra_path, extra_surface in zip(extras_paths, def_extras):
name = extra_path.split('/')[-1].split(
f'{source_block}')[-1].replace('.', '_deformable.')
# if '_stitched' in name:
# name = name.replace('_stitched', '')
if not os.path.exists(f'{out_path}{name}') or rerun:
io.WriteOBJ(extra_surface.vertices, extra_surface.indices,
f'{out_path}{name}')
print('Done registering head blocks to middle block.')
def exvivo_to_blocks(stacked_blocks_dir):
# This is registered from blocks to exvivo, so phi is needed to bring the exvivo MR image to the block images
# Need to determine the grid to sample the MR onto
rabbit_dir = f'/hdscratch/ucair/{rabbit}/blockface/'
block_list = sorted(glob.glob(f'{rabbit_dir}block*'))
orig_dir = f'/home/sci/blakez/ucair/{rabbit}/rawVolumes/ExVivo_2018-07-26/'
spacing = []
origin = []
size = []
for block_path in block_list:
if 'block07' in block_path:
hdr = tools.read_mhd_header(f'{block_path}/volumes/raw/difference_volume.mhd')
else:
hdr = tools.read_mhd_header(f'{block_path}/volumes/deformable/difference_volume_deformable.mhd')
spacing.append(np.array([float(x) for x in hdr['ElementSpacing'].split(' ')]))
origin.append(np.array([float(x) for x in hdr['Offset'].split(' ')]))
size.append(np.array([float(x) for x in hdr['DimSize'].split(' ')]))
spacing = np.stack(spacing)
origin = np.stack(origin)
size = np.stack(size)
extent = size * spacing + origin
aff_grid_size = np.array((512, 512, 512))
aff_grid_origin = np.min(origin, axis=0)
aff_grid_spacing = (np.max(extent, axis=0) - aff_grid_origin) / aff_grid_size
aff_grid = core.StructuredGrid(
size=aff_grid_size.tolist()[::-1],
spacing=aff_grid_spacing.tolist()[::-1],
origin=aff_grid_origin.tolist()[::-1],
device=device,
channels=3
)
aff_grid.set_size(size=(512, 512, 512), inplace=True)
aff_grid.set_to_identity_lut_()
# Load the affine
aff = np.loadtxt(f'{orig_dir}blocks_to_exvivo_affine.txt')
aff = torch.tensor(aff, device=device, dtype=torch.float32)
# Apply the FORWARD affine to the deformation
# aff = aff.inverse()
a = aff[0:3, 0:3].float()
t = aff[-0:3, 3].float()
# Create a deformation from the affine that lives in the stacked blocks space
aff_grid.data = aff_grid.data.flip(0)
aff_grid.data = torch.matmul(a, aff_grid.data.permute(list(range(1, 3 + 1)) + [0]).unsqueeze(-1))
aff_grid.data = (aff_grid.data.squeeze() + t).permute([-1] + list(range(0, 3)))
aff_grid.data = aff_grid.data.flip(0)
# Load the defromabale transformation
phi = io.LoadITKFile(
f'{orig_dir}blocks_phi_to_exvivo.mhd', device=device
)
phi.set_size((256, 256, 256))
# Compose the grids
deformation = so.ComposeGrids.Create(device=device)([aff_grid, phi])
return deformation
def corrections(rabbit_dir, block, corr_num):
block_path = f'{rabbit_dir}{block}/'
image_list = sorted(glob.glob(f'{block_path}/images/filtered/*'))
with open(f'{block_path}/volumes/raw/{block}_affine_dictionary.yaml',
'r') as f:
Adict = yaml.load(f, Loader=yaml.FullLoader)
surface_list = [x for x in image_list if 'surface' in x]
scatter_list = [x for x in image_list if 'scatter' in x]
# Load the surface image
ImSurface_good = io.LoadITKFile(surface_list[corr_num - 1], device=device)
ImSurface_good.set_spacing_(Adict['spacing'])
ImSurface_good.set_origin_(
-1 * (ImSurface_good.size * ImSurface_good.spacing) / 2)
ImSurface_good /= 255.0
ImSurface_bad = io.LoadITKFile(surface_list[corr_num], device=device)
ImSurface_bad.set_spacing_(Adict['spacing'])
ImSurface_bad.set_origin_(-1 *
(ImSurface_bad.size * ImSurface_bad.spacing) / 2)
ImSurface_bad /= 255.0
# Apply the affines to both of the images
aff_filter = so.AffineTransform.Create(affine=torch.tensor(Adict[corr_num -
1],
device=device),
device=device)
ImSurface_good = aff_filter(ImSurface_good)
aff_filter = so.AffineTransform.Create(affine=torch.tensor(Adict[corr_num],
device=device),
device=device)
ImSurface_bad = aff_filter(ImSurface_bad)
points = torch.tensor(tools.LandmarkPicker(
[ImSurface_bad[0].squeeze().cpu(), ImSurface_good[0].squeeze().cpu()]),
dtype=torch.float32,
device=device).permute(1, 0, 2)
# Change to real coordinates
points *= torch.cat([
ImSurface_good.spacing[None, None, :], ImSurface_good.spacing[None,
None, :]
], 0)
points += torch.cat([
ImSurface_good.origin[None, None, :], ImSurface_good.origin[None,
None, :]
], 0)
aff_filter = so.AffineTransform.Create(points[1], points[0], device=device)
corr_affine = torch.eye(3, device=device, dtype=torch.float32)
corr_affine[0:2, 0:2] = aff_filter.affine
corr_affine[0:2, 2] = aff_filter.translation
ImPrev = ImSurface_good.copy()
for scat, surf in zip(scatter_list[corr_num:], surface_list[corr_num:]):
print(f'Registering {scat.split("/")[-1]} .... ')
sys.stdout.flush()
image_num = scat.split('/')[-1].split('_')[1]
# Get the number the file is from the start
dist = scatter_list.index(scat)
# Load the next image
ImScatter = io.LoadITKFile(scat, device=device)
ImScatter.set_spacing_(ImPrev.spacing)
ImScatter.set_origin_(ImPrev.origin)
ImScatter /= 255.0
ImSurface = io.LoadITKFile(surf, device=device)
ImSurface.set_spacing_(ImPrev.spacing)
ImSurface.set_origin_(ImPrev.origin)
ImSurface /= 255.0
difference = ImScatter - ImSurface
ImDifference = core.StructuredGrid(
ImSurface.shape()[1:],
tensor=difference.data[2].unsqueeze(0),
spacing=ImSurface.spacing,
origin=ImSurface.origin,
device=device,
dtype=torch.float32,
channels=1)
affine = torch.mm(corr_affine, torch.tensor(Adict[dist],
device=device))
# Save out the images
aff_filter = so.AffineTransform.Create(affine=affine, device=device)
aff_scatter = aff_filter(ImScatter)
aff_surface = aff_filter(ImSurface)
aff_difference = aff_filter(ImDifference)
# difference = (difference - difference.min()) / (difference.max() - difference.min())
io.SaveITKFile(
aff_scatter,
f'{block_path}/volumes/raw/scatter/IMG_{image_num}_scatter.mhd')
io.SaveITKFile(
aff_surface,
f'{block_path}/volumes/raw/surface/IMG_{image_num}_surface.mhd')
io.SaveITKFile(
aff_difference,
f'{block_path}/volumes/raw/difference/IMG_{image_num}_difference.mhd'
)
Adict[dist] = affine.detach().cpu().clone().tolist()
ImPrev = aff_scatter.copy()
with open(f'{block_path}/volumes/raw/{block}_affine_dictionary.yaml',
'w') as f:
yaml.dump(Adict, f)
def SolveAffineGrid(source_image, input_affine):
'''Takes a source volume and an affine matrix and solves for the necessary
grid in the target space of the affine. Essentially calculates the bounding
box of the source image after apply the affine transformation
source_image = volume to be transformed by the affine (Image3D type)
input_affine = the affine transformation (4x4)
Returns the grid for the transformed source volume in real coordinates'''
# Make sure we don't mess with the incoming affine
affine = np.copy(input_affine)
# Get parameters from the incoming source image
in_sz = source_image.size.tolist()
in_sp = source_image.spacing.tolist()
in_or = source_image.origin.tolist()
# Extract the pure rotation and ignore scaling to find the final size of the volume
U, s, V = np.linalg.svd(affine[0:3, 0:3])
rotMat = np.eye(4)
rotMat[0:3, 0:3] = np.dot(U, V)
# Get the corners of the volume in index coordinates
inputCorners = np.array([0, 0, 0, 1])
inputCorners = np.vstack((inputCorners, np.array([in_sz[0], 0, 0, 1])))
inputCorners = np.vstack((inputCorners, np.array([0, in_sz[1], 0, 1])))
inputCorners = np.vstack((inputCorners, np.array([in_sz[0], in_sz[1], 0, 1])))
# Account for the case that the source image is 3D
if len(source_image.size) == 3:
inputCorners = np.vstack((inputCorners, np.array([0, 0, in_sz[2], 1])))
inputCorners = np.vstack((inputCorners, np.array([in_sz[0], 0, in_sz[2], 1])))
inputCorners = np.vstack((inputCorners, np.array([0, in_sz[1], in_sz[2], 1])))
inputCorners = np.vstack((inputCorners, np.array([in_sz[0], in_sz[1], in_sz[2], 1])))
# Define the index corners to find the final size of the transformed volume
indexCorners = np.matrix(inputCorners)
# Find the real corners of the input volume for finding the output origin
realCorners = np.matrix(np.multiply(inputCorners, np.array(in_sp + [1])) + np.array(in_or + [0]))
# Apply the transformations to the real and index corners
# Need to subtract the mean and apply around the center
# Translation is not going to effect the index size, but will effect the origin of the block for real coordinates
outRealCorners = np.matmul(affine, realCorners.transpose())
outIndxCorners = np.matmul(rotMat, indexCorners.transpose())
# Find the size in real and index coordinates of the output volume
realSize = (np.max(outRealCorners, 1) - np.min(outRealCorners, 1))[0:3]
indexSize = (np.max(outIndxCorners, 1) - np.min(outIndxCorners, 1))[0:3]
# The index size is the size of the output volume
out_sz = np.squeeze(np.array(np.ceil(indexSize).astype(int)))
out_sz[out_sz == 0] = 1
# We can divide index size into real size to get the spacing of the real volume; Need to account for 2D zero
out_sp = np.squeeze(np.array(np.divide(realSize, indexSize, where=indexSize != 0)))
out_sp[out_sp == 0] = 1
# Find the output origin by taking the min in each dimension of the real transformed corners
out_or = np.squeeze(np.array(np.min(outRealCorners, 1)))[0:3]
# Make the output structured grid
out_tensor = core.StructuredGrid(
size=out_sz.copy(),
origin=out_or.copy(),
spacing=out_sp.copy(),
device=source_image.device,
channels=1
)
return out_tensor
def front_face_stacking(rabbit, base_dir='/hdscratch/ucair/'):
rabbit_dir = f'{base_dir}{rabbit}/blockface/'
raw_ext = '/surfaces/raw/'
# rigid_ext = '/surfaces/rigid/'
ff_ext = '/surfaces/frontface/'
deform_ext = '/surfaces/deformable/'
vol_ext = '/volumes/raw/'
stitch_ext = '/surfaces/raw/stitching/deformable/'
# Get a list of the blocks
block_list = sorted(glob.glob(f'{rabbit_dir}block*'))
# Determine the middle block
middle_block = block_list[4]
foot_blocks = block_list[block_list.index(middle_block):]
head_blocks = block_list[:block_list.index(middle_block) + 1][::-1]
rerun = True
# skip_blocks = ['block05', 'block06', 'block07', 'block08', 'block09', 'block10', 'block11', 'block12']
skip_blocks = []
if rerun:
mid_block = middle_block.split('/')[-1]
if not os.path.exists(f'{rabbit_dir}{mid_block}{ff_ext}'):
os.makedirs(f'{rabbit_dir}{mid_block}{ff_ext}')
affine_tform = torch.eye(4)
np.savetxt(
f'{rabbit_dir}{mid_block}{raw_ext}{mid_block}_front_face_tform.txt',
affine_tform.numpy())
# Copy the files from raw to deformable for the middle surface
if os.path.exists(f'{rabbit_dir}{mid_block}{stitch_ext}'):
mid_path = f'{mid_block}{stitch_ext}'
else:
mid_path = f'{mid_block}{raw_ext}'
files = [
f'{rabbit_dir}{mid_path}{mid_block}_decimate.obj',
f'{rabbit_dir}{mid_path}{mid_block}_ext.obj',
]
if os.path.exists(f'{rabbit_dir}{mid_path}{mid_block}_foot.obj'):
files += [f'{rabbit_dir}{mid_path}{mid_block}_foot.obj']
if os.path.exists(f'{rabbit_dir}{mid_path}{mid_block}_head.obj'):
files += [f'{rabbit_dir}{mid_path}{mid_block}_head.obj']
if os.path.exists(
f'{rabbit_dir}{mid_path}{mid_block}_foot_support.obj'):
files += [f'{rabbit_dir}{mid_path}{mid_block}_foot_support.obj']
if os.path.exists(
f'{rabbit_dir}{mid_path}{mid_block}_head_support.obj'):
files += [f'{rabbit_dir}{mid_path}{mid_block}_head_support.obj']
out_names = []
for path in files:
name = path.split('/')[-1].split(f'{mid_block}')[-1].replace(
'.', '_front_face.')
out_path = f'{rabbit_dir}{mid_block}{ff_ext}{mid_block}'
out_names += [f'{out_path}{name}']
for in_file, out_file in zip(files, out_names):
shutil.copy(in_file, out_file)
# Loop over the foot blocks
for i, block_path in enumerate(foot_blocks, 1):
if i == len(foot_blocks):
break
target_block = block_path.split('/')[-1]
source_block = foot_blocks[i].split('/')[-1]
if source_block in skip_blocks:
continue
if os.path.exists(f'{rabbit_dir}{source_block}{stitch_ext}'):
mid_path = f'{source_block}{stitch_ext}'
else:
mid_path = f'{source_block}{raw_ext}'
target_surface_path = f'{rabbit_dir}{target_block}{ff_ext}{target_block}_foot_front_face.obj'
source_surface_path = f'{rabbit_dir}{mid_path}{source_block}_head.obj'
extras_paths = [
f'{rabbit_dir}{mid_path}{source_block}_decimate.obj',
f'{rabbit_dir}{mid_path}{source_block}_ext.obj',
]
if i < len(foot_blocks) - 1:
extras_paths += [f'{rabbit_dir}{mid_path}{source_block}_foot.obj']
if os.path.exists(
f'{rabbit_dir}{mid_path}{source_block}_foot_support.obj'):
extras_paths += [
f'{rabbit_dir}{mid_path}{source_block}_foot_support.obj'
]
if os.path.exists(
f'{rabbit_dir}{source_block}{ff_ext}{source_block}_head_front_face.obj'
) and not rerun:
print(
f'The front face surface for {source_block} already exists ... Next block'
)
continue
try:
verts, faces = io.ReadOBJ(target_surface_path)
tar_surface = core.TriangleMesh(verts, faces)
tar_surface.to_(device)
except IOError:
print(
f'The front face foot surface for {target_block} was not found ... Next block'
)
continue
# Need to see if the target needs any support
support_block = block_list[block_list.index(block_path) -
1].split('/')[-1]
if os.path.exists(
f'{rabbit_dir}{support_block}{ff_ext}{support_block}_foot_support_front_face.obj'
):
verts, faces = io.ReadOBJ(
f'{rabbit_dir}{support_block}{ff_ext}{support_block}_foot_support_front_face.obj'
)
tar_surface.add_surface_(verts.to(device=device),
faces.to(device=device))
try:
verts, faces = io.ReadOBJ(source_surface_path)
src_surface = core.TriangleMesh(verts, faces)
src_surface.to_(device)
src_surface.flip_normals_()
except IOError:
print(
f'The raw head surface for {source_block} was not found ... Next block'
)
continue
extra_surfaces = []
for path in extras_paths:
try:
verts, faces = io.ReadOBJ(path)
except IOError:
extra_name = path.split('/')[-1]
print(
f'{extra_name} not found as an extra ... removing from list'
)
_ = extras_paths.pop(extras_paths.index(path))
continue
extra_surfaces += [core.TriangleMesh(verts, faces)]
extra_surfaces[-1].to_(device)
# Load or create the dictionary for registration
try:
with open(
f'{rabbit_dir}{source_block}{raw_ext}{source_block}_front_face_config.yaml',
'r') as f:
params = yaml.load(f, Loader=yaml.FullLoader)
except IOError:
params = {
'spatial_sigma': [2.0, 1.0],
'affine_lr': 1.0e-06,
'translation_lr': 1.0e-04,
'converge': 0.01,
'rigid_transform': True
}
print(f'Registering {source_block} to {target_block}:')
affine_tform = front_face_register(
tar_surface.copy(),
src_surface.copy(),
spatial_sigma=params['spatial_sigma'],
affine_lr=params['affine_lr'],
translation_lr=params['translation_lr'],
converge=params['converge'],
device=device)
# Apply the affine to the source element and the excess
aff_tformer = uo.AffineTransformSurface.Create(affine_tform,
device=device)
aff_src_surface = aff_tformer(src_surface)
aff_extra_surface = []
for surface in extra_surfaces:
aff_extra_surface += [aff_tformer(surface)]
out_path = f'{rabbit_dir}{source_block}{ff_ext}{source_block}'
if not os.path.exists(f'{rabbit_dir}{source_block}{ff_ext}'):
os.makedirs(f'{rabbit_dir}{source_block}{ff_ext}')
if not os.path.exists(f'{out_path}_head_front_face.obj') or rerun:
io.WriteOBJ(aff_src_surface.vertices, aff_src_surface.indices,
f'{out_path}_head_front_face.obj')
for extra_path, extra_surface in zip(extras_paths, aff_extra_surface):
name = extra_path.split('/')[-1].split(
f'{source_block}')[-1].replace('.', '_front_face.')
if not os.path.exists(f'{out_path}{name}') or rerun:
io.WriteOBJ(extra_surface.vertices, extra_surface.indices,
f'{out_path}{name}')
# Save out the parameters:
with open(
f'{rabbit_dir}{source_block}{raw_ext}{source_block}_front_face_config.yaml',
'w') as f:
yaml.dump(params, f)
# Save the affine in the volumes and in the surfaces location
np.savetxt(
f'{rabbit_dir}{source_block}{raw_ext}{source_block}_front_face_tform.txt',
affine_tform.numpy())
print('Done registering foots blocks to middle block.')
# Loop over the head blocks
for i, block_path in enumerate(head_blocks, 1):
if i == len(head_blocks):
break
target_block = block_path.split('/')[-1]
source_block = head_blocks[i].split('/')[-1]
if source_block in skip_blocks:
continue
if os.path.exists(f'{rabbit_dir}{source_block}{stitch_ext}'):
mid_path = f'{source_block}{stitch_ext}'
else:
mid_path = f'{source_block}{raw_ext}'
target_surface_path = f'{rabbit_dir}{target_block}{ff_ext}{target_block}_head_front_face.obj'
source_surface_path = f'{rabbit_dir}{mid_path}{source_block}_foot.obj'
extras_paths = [
f'{rabbit_dir}{mid_path}{source_block}_decimate.obj',
f'{rabbit_dir}{mid_path}{source_block}_ext.obj'
]
if i < len(head_blocks) - 1:
extras_paths += [f'{rabbit_dir}{mid_path}{source_block}_head.obj']
if os.path.exists(
f'{rabbit_dir}{mid_path}{source_block}_head_support.obj'):
extras_paths += [
f'{rabbit_dir}{mid_path}{source_block}_head_support.obj'
]
if os.path.exists(
f'{rabbit_dir}{source_block}{ff_ext}{source_block}_foot_front_face.obj'
) and not rerun:
print(
f'The front face surface for {source_block} already exists ... Next block'
)
continue
try:
verts, faces = io.ReadOBJ(target_surface_path)
tar_surface = core.TriangleMesh(verts, faces)
tar_surface.to_(device)
except IOError:
print(
f'The front face foot surface for {target_block} was not found ... Next block'
)
continue
# Need to see if the target needs any support
support_block = block_list[block_list.index(block_path) +
1].split('/')[-1]
if os.path.exists(
f'{rabbit_dir}{support_block}{ff_ext}{support_block}_head_support_front_face.obj'
):
verts, faces = io.ReadOBJ(
f'{rabbit_dir}{support_block}{ff_ext}{support_block}_head_support_front_face.obj'
)
tar_surface.add_surface_(verts.to(device=device),
faces.to(device=device))
try:
verts, faces = io.ReadOBJ(source_surface_path)
src_surface = core.TriangleMesh(verts, faces)
src_surface.to_(device)
src_surface.flip_normals_()
except IOError:
print(
f'The raw foot surface for {source_block} was not found ... Next block'
)
continue
extra_surfaces = []
for path in extras_paths:
try:
verts, faces = io.ReadOBJ(path)
except IOError:
extra_name = path.split('/')[-1]
print(
f'{extra_name} not found as an extra ... removing from list'
)
_ = extras_paths.pop(extras_paths.index(path))
continue
extra_surfaces += [core.TriangleMesh(verts, faces)]
extra_surfaces[-1].to_(device)
# Load or create the dictionary for registration
try:
with open(
f'{rabbit_dir}{source_block}{raw_ext}{source_block}_front_face_config.yaml',
'r') as f:
params = yaml.load(f, Loader=yaml.FullLoader)
except IOError:
params = {
'spatial_sigma': [2.0, 1.0],
'affine_lr': 1.0e-06,
'translation_lr': 1.0e-04,
'converge': 0.01,
'rigid_transform': True
}
print(f'Registering {source_block} to {target_block}:')
affine_tform = front_face_register(
tar_surface.copy(),
src_surface.copy(),
spatial_sigma=params['spatial_sigma'],
affine_lr=params['affine_lr'],
translation_lr=params['translation_lr'],
converge=params['converge'],
device=device)
# Apply the affine to the source element and the excess
aff_tformer = uo.AffineTransformSurface.Create(affine_tform,
device=device)
aff_src_surface = aff_tformer(src_surface)
aff_extra_surface = []
for surface in extra_surfaces:
aff_extra_surface += [aff_tformer(surface)]
out_path = f'{rabbit_dir}{source_block}{ff_ext}{source_block}'
if not os.path.exists(f'{rabbit_dir}{source_block}{ff_ext}'):
os.makedirs(f'{rabbit_dir}{source_block}{ff_ext}')
if not os.path.exists(f'{out_path}_foot_front_face.obj') or rerun:
io.WriteOBJ(aff_src_surface.vertices, aff_src_surface.indices,
f'{out_path}_foot_front_face.obj')
for extra_path, extra_surface in zip(extras_paths, aff_extra_surface):
name = extra_path.split('/')[-1].split(
f'{source_block}')[-1].replace('.', '_front_face.')
if not os.path.exists(f'{out_path}{name}') or rerun:
io.WriteOBJ(extra_surface.vertices, extra_surface.indices,
f'{out_path}{name}')
# Save out the parameters:
with open(
f'{rabbit_dir}{source_block}{raw_ext}{source_block}_front_face_config.yaml',
'w') as f:
yaml.dump(params, f)
# Save the affine in the volumes and in the surfaces location
np.savetxt(
f'{rabbit_dir}{source_block}{raw_ext}{source_block}_front_face_tform.txt',
affine_tform.numpy())
print('Done registering head blocks to middle block.')
def process_mic(rabbit):
raw_mic_dir = f'/hdscratch/ucair/{rabbit}/microscopic/'
bf_dir = f'/hdscratch/ucair/{rabbit}/blockface/'
raw_bf_dir = f'/hdscratch/ucair/blockface/{rabbit}/'
block_list = sorted(glob.glob(f'{raw_mic_dir}/*'))
# for block_path in block_list:
# block = block_path.split('/')[-1]
#
# mic_list = sorted(glob.glob(f'{block_path}/raw/*_image.tif'))
#
# img_nums = [x.split('/')[-1].split('_')[1] for x in mic_list]
#
# # Load the image
# for img in img_nums:
# print(f'Processing {block}, {img} ... ', end='')
# mic_file = f'{raw_mic_dir}{block}/hdf5/{block}_img{img}_image.hdf5'
#
# mic = io.LoadITKFile(f'{raw_mic_dir}{block}/raw/IMG_{img}_histopathology_image.tif')
#
# with h5py.File(mic_file, 'w') as f:
# g = f.create_group('RawImage')
# g.create_dataset('ImageData', data=mic.data.numpy())
#
# print('Done')
for block_path in block_list:
block = block_path.split('/')[-1]
mic_list = sorted(glob.glob(f'{block_path}/raw/*_image.tif'))
img_nums = [x.split('/')[-1].split('_')[1] for x in mic_list]
for img in img_nums:
mic_file = f'{raw_mic_dir}{block}/hdf5/{block}_img{img}_image.hdf5'
mic_seg = f'{raw_mic_dir}{block}/hdf5/{block}_img{img}_label.hdf5'
blockface_image = f'{bf_dir}{block}/volumes/raw/difference/IMG_{img}_difference.mhd'
blockface_label = f'{bf_dir}{block}/volumes/raw/segmentation/IMG_{img}_segmentation.mhd'
meta_dict = {}
with h5py.File(mic_file, 'r') as f:
for key in f['RawImage'].attrs:
meta_dict[key] = f['RawImage'].attrs[key]
if 'Affine' in meta_dict.keys():
continue
blockface = io.LoadITKFile(blockface_image, device=device)
label_data = io.LoadITKFile(blockface_label, device=device)
blockface = (blockface - blockface.min()) / (blockface.max() -
blockface.min())
label = blockface.clone()
label.data = label_data.data.clone()
print(f'Affine Registering ... ')
aff_mic, aff_mic_seg, affine = tools.process_mic(mic_file,
mic_seg,
blockface,
label,
device=device)
print(f'Done')
aff_mic *= aff_mic_seg
blockface *= label
blockface_s = core.StructuredGrid.FromGrid(
blockface, tensor=blockface[0].unsqueeze(0), channels=1)
aff_mic = (aff_mic - aff_mic.min()) / (aff_mic.max() -
aff_mic.min())
print(f'Deformable Registering Labels ... ')
def_label, label_deformation = match_bf_mic(aff_mic_seg,
label,
steps=[0.01, 0.005],
scales=[4, 1],
gauss=True)
print(f'Done')
label_def_mic = so.ApplyGrid(label_deformation,
device=device)(aff_mic,
label_deformation)
print(f'Deformable Registering Images ... ')
def_image, image_deformation = match_bf_mic(
label_def_mic,
blockface_s,
steps=[0.01, 0.01],
scales=[2, 1],
)
print(f'Done')
composer = so.ComposeGrids(device=device,
dtype=torch.float32,
padding_mode='border')
deformation = composer([image_deformation, label_deformation])
def_mic = so.ApplyGrid(deformation, device=device)(aff_mic,
deformation)
try:
with h5py.File(mic_file, 'r') as f:
mic = f['ImageData'][:]
except KeyError:
with h5py.File(mic_file, 'r') as f:
mic = f['RawImage/ImageData'][:]
mic = core.StructuredGrid(mic.shape[1:],
tensor=torch.tensor(mic,
dtype=torch.float32,
device=device),
device=device,
dtype=torch.float32,
channels=3)
print('Saving ... ')
with h5py.File(mic_file, 'w') as f:
g = f.create_group('RawImage')
d = f.create_group('Deformation')
g.create_dataset('ImageData', data=mic.data.cpu().numpy())
d.create_dataset('Phi', data=deformation.data.cpu().numpy())
g.attrs['Shape'] = list(mic.shape())
g.attrs['Spacing'] = mic.spacing.tolist()
g.attrs['Origin'] = mic.origin.tolist()
g.attrs['Affine'] = affine.tolist()
d.attrs['Shape'] = list(deformation.shape())
d.attrs['Spacing'] = deformation.spacing.tolist()
d.attrs['Origin'] = deformation.origin.tolist()
io.SaveITKFile(
def_mic,
f'{raw_mic_dir}{block}/volume/images/IMG_{img}_def_histopathology.mhd'
)
print('Done')
plt.close('all')
print('All Done')
def tempRecon(Obj, opt):
if not os.path.exists(f'{Obj.rabbitDirectory}/rawVolumes/TemperatureRecon/'):
os.makedirs(f'{Obj.rabbitDirectory}/rawVolumes/TemperatureRecon/')
size, orgn = _convert_mat_files(Obj)
unionOrigin = np.min(orgn, 0).tolist()
adj = np.abs(np.max(orgn, 0) - np.min(orgn, 0))
unionSize = (np.max(size, 0) + adj).tolist()
unionSize = [int(x) for x in unionSize]
# We are going to create a config file for each registration
sonicationList = natural_sort(glob.glob(Obj.rabbitDirectory + '/rawVolumes/TemperatureRecon/*'))
# Need to remove the resample from the list
sonicationList = [x for x in sonicationList if '_resample' not in x]
sonicationList = [x for x in sonicationList if 'RegVols' not in x]
lastName = sonicationList[-1].split('/')[-1]
preLength = len(lastName.split('_')[0])
lastName = lastName[:preLength] + '_----' + lastName[preLength:]
# Set up the Config directory
if not os.path.exists(Obj.rabbitDirectory + '/ConfigFiles/temperatureConfigs/'):
os.makedirs(Obj.rabbitDirectory + '/ConfigFiles/temperatureConfigs/')
# Construct teh common Grid: at this point assuming isotropic spacing of 1
if 'unionSize' in locals():
unionSpacing = [1.0, 1.0, 1.0]
unionIm = core.StructuredGrid(
size=unionSize,
origin=unionOrigin,
spacing=unionSpacing,
channels=1,
device=device
)
def_im = unionIm.clone()
else:
configList = glob.glob(Obj.rabbitDirectory + '/ConfigFiles/temperatureConfigs/*')
tmpObj = Config.Load(TempConfigSpec, configList[0])
unionSize = tmpObj.unionSize
unionOrigin = tmpObj.unionOrigin
unionSpacing = tmpObj.unionSpacing
unionIm = core.StructuredGrid(
size=unionSize,
origin=unionOrigin,
spacing=unionSpacing,
channels=1,
device=device
)
def_im = unionIm.clone()
# We need to register the last image to the VIBE
son = sonicationList[-1]
try:
regObj = Config.Load(TempConfigSpec,
Obj.rabbitDirectory + '/ConfigFiles/temperatureConfigs/{0}_config.yaml'.format(
son.split('/')[-1]))
except IOError:
regObj = Config.SpecToConfig(TempConfigSpec)
regObj.Niter = [40]
regObj.gaussSigma = [0.1]
regObj.gaussKernel = [3]
regObj.regWeight = [2.0]
regObj.stepSize = [0.004]
regObj.I0 = 'None'
regObj.I1 = 'None'
regObj.unionSize = unionSize
regObj.unionSpacing = unionSpacing
regObj.unionOrigin = unionOrigin
# Need to get the name of the volume
outName = son.split('/')[-1]
preLength = len(outName.split('_')[0])
outName = outName[:preLength] + '_----' + outName[preLength:]
regObj.outputPath = Obj.rabbitDirectory + '/elastVolumes/TemperatureRecon/' + son.split('/')[-1]
if not os.path.exists(regObj.outputPath):
os.makedirs(regObj.outputPath)
plt.close('all')
# Get the deformation field that takes the first of the sonication images to the final
h_post = ER.MultiscaleElast(regObj, opt)
io.SaveITKFile(h_post, regObj.outputPath + '/final_deformation_Is_to_It.mha')
rc.WriteConfig(TempConfigSpec, regObj,
Obj.rabbitDirectory + '/ConfigFiles/temperatureConfigs/{0}_config.yaml'.format(son.split('/')[-1]))
print('Applying elast transformation to {0} volumes .... '.format(son.split('/')[-1]), end='')
sys.stdout.flush()
for image in glob.glob('{0}/rawVolumes/TemperatureRecon/'.format(Obj.rabbitDirectory) + son.split('/')[-1] + '/*'):
# Update the name
outName = image.split('/')[-1]
outName = outName[:preLength + 4] + 'e' + outName[preLength + 5:]
im = io.LoadITKFile(image, device=device)
temp = so.ResampleWorld.Create(unionIm, device=device)(im)
def_im = so.ApplyGrid.Create(h_post, device=device)(temp)
io.SaveITKFile(def_im, regObj.outputPath + '/' + outName)
# common.DebugHere()
# # Need to resample the final image onto the new grid
# for image in glob.glob('{0}/rawVolumes/TemperatureRecon/'.format(Obj.rabbitDirectory) + sonicationList[-1].split('/')[-1] + '/*'):
# outputPath = Obj.rabbitDirectory + '/rawVolumes/TemperatureRecon/' + sonicationList[-1].split('/')[-1] + '_resample'
# if not os.path.exists(outputPath):
# os.makedirs(outputPath)
# outName = image.split('/')[-1]
# im = common.LoadITKImage(image, ca.MEM_DEVICE)
# cc.ResampleWorld(unionIm, im, bg=3)
# common.SaveITKImage(unionIm, outputPath + '/' + outName)
for son in sonicationList[:-1]:
# Create the registration object for elastic
try:
regObj = Config.Load(TempConfigSpec,
Obj.rabbitDirectory + '/ConfigFiles/temperatureConfigs/{0}_config.yaml'.format(
son.split('/')[-1]))
except IOError:
regObj = Config.SpecToConfig(TempConfigSpec)
# Need to get the name of the volume
outName = son.split('/')[-1]
preLength = len(outName.split('_')[0])
outName = outName[:preLength] + '_----' + outName[preLength:]
# Always set the target image to the last file - have to get the first mag image!
regObj.I0 = son + '/../RegVols/' + outName[0:5] + '.nii.gz'
# Do we have to get the file that has been registered to the vibe??
regObj.I1 = sonicationList[-1] + '/../RegVols/' + lastName[0:5] + '.nii.gz'
# Set the grid parameters of the unionFOV
regObj.unionSize = unionSize
regObj.unionSpacing = unionSpacing
regObj.unionOrigin = unionOrigin
# if sonicationList.index(son) < 8:
# regObj.stepSize = [0.000]
# else:
# regObj.stepSize = [0.004]
regObj.outputPath = Obj.rabbitDirectory + '/elastVolumes/TemperatureRecon/' + son.split('/')[-1]
if not os.path.exists(regObj.outputPath):
os.makedirs(regObj.outputPath)
plt.close('all')
# Get the deformation field that takes the first of the sonication images to the final
h = ER.MultiscaleElast(regObj, opt)
# Need to compose the h field to get the full deformation
h_comp = so.ComposeGrids.Create(device=device)([h, h_post])
io.SaveITKFile(h_comp, regObj.outputPath + '/final_deformation_Is_to_It.mha')
# Write out the config file
if not os.path.exists(Obj.rabbitDirectory + '/ConfigFiles/temperatureConfigs'):
os.makedirs(Obj.rabbitDirectory + '/ConfigFiles/temperatureConfigs')
rc.WriteConfig(TempConfigSpec, regObj,
Obj.rabbitDirectory + '/ConfigFiles/temperatureConfigs/{0}_config.yaml'.format(
son.split('/')[-1]))
# Need to itterate through all the mag and phase images and apply the deformation
print('Applying elast transformation to {0} volumes .... '.format(son.split('/')[-1]), end='')
sys.stdout.flush()
for image in glob.glob(
'{0}/rawVolumes/TemperatureRecon/'.format(Obj.rabbitDirectory) + son.split('/')[-1] + '/*'):
# Update the name
outName = image.split('/')[-1]
outName = outName[:preLength + 4] + 'e' + outName[preLength + 5:]
im = io.LoadITKFile(image, device=device)
temp = so.ResampleWorld.Create(unionIm, device=device)(im)
def_im = so.ApplyGrid.Create(h_post, device=device)(temp)
io.SaveITKFile(def_im, regObj.outputPath + '/' + outName)
def eval_R1(rabbit_list):
def_diff = []
aff_diff = []
print('Eval R1:')
for r in rabbit_list:
print(f'Processing Rabbit {r} ... ', end='')
in_path = f'/hdscratch/ucair/{r}/'
if r == '18_047':
in_path = f'/hdscratch2/{r}/'
block_list = sorted(glob.glob(f'{in_path}/blockface/block*'))
for b_path in block_list:
b = b_path.split('/')[-1]
lm_list = sorted(
glob.glob(f'{in_path}/microscopic/{b}/landmarks/R1*'))
if not lm_list:
continue
for lm in lm_list:
i = lm.split('_')[6]
r_img = io.LoadITKFile(
f'{in_path}microscopic/{b}/segmentations/IMG_{i}/img_{i}_red.nii.gz',
device=device)
def_dir = f'{in_path}microscopic/{b}/deformations/'
# Load the affine
aff = np.loadtxt(
glob.glob(f'{def_dir}img_{i}_affine_to_blockface.txt')[0])
aff = torch.tensor(aff, device=device, dtype=torch.float32)
# Load the deformation
deformation_data = io.LoadITKFile(
f'{def_dir}/img_{i}_deformation_to_blockface.mhd',
device=device)
deformation = core.StructuredGrid(
size=deformation_data.size[0:2],
spacing=deformation_data.spacing[1:3],
origin=deformation_data.origin[1:3],
device=deformation_data.device,
tensor=deformation_data.data.squeeze().permute(2, 0, 1),
channels=2)
# Apply the inverse affine to the grid
aff = aff.inverse()
a = aff[0:2, 0:2].float()
t = aff[-0:2, 2].float()
aff_deformation = deformation.clone()
aff_deformation.set_to_identity_lut_()
deformation.data = torch.matmul(
a.unsqueeze(0).unsqueeze(0),
deformation.data.permute(list(range(1, 2 + 1)) +
[0]).unsqueeze(-1))
deformation.data = (deformation.data.squeeze() +
t).permute([-1] + list(range(0, 2)))
aff_deformation.data = torch.matmul(
a.unsqueeze(0).unsqueeze(0),
aff_deformation.data.permute(list(range(1, 2 + 1)) +
[0]).unsqueeze(-1))
aff_deformation.data = (aff_deformation.data.squeeze() +
t).permute([-1] + list(range(0, 2)))
in_bf = f'{in_path}blockface/{b}/landmarks/'
bf_lm = [
torch.tensor(np.loadtxt(
f'{in_bf}/R1_{r}_{b}_IMG_{i}_blockface_landmarks.txt'),
device=device)
]
in_mic = f'{in_path}microscopic/{b}/landmarks/'
mic_lm = torch.tensor(np.loadtxt(
f'{in_mic}/R1_{r}_{b}_IMG_{i}_microscopic_landmarks.txt'),
device=device)
def_point = sample_points(deformation, bf_lm)
aff_point = sample_points(aff_deformation, bf_lm)
# idx_def = def_point[0]/ r_img.spacing
aff_diff.append((mic_lm - aff_point) * 0.00176)
def_diff.append((mic_lm - def_point) *
0.00176) # spacing from histology in mm
print('done')
def_dist = torch.sqrt((torch.cat(def_diff, 0)**2).sum(-1))
aff_dist = torch.sqrt((torch.cat(aff_diff, 0)**2).sum(-1))
print('Eval R1: Done')
return def_dist, aff_dist
def example():
f = box_image(
core.StructuredGrid((256, 256), spacing=[1.0, 1.0], device=device))
d = box_image(core.StructuredGrid(
(256, 256), spacing=[1.0, 1.0], device=device),
break_width=50,
val=500.0) + 0.1
m = box_image(core.StructuredGrid((256, 256),
spacing=[1.0, 1.0],
device=device),
break_width=50,
val=1.0)
f = so.Gaussian.Create(channels=1,
kernel_size=5,
sigma=2,
dim=2,
device=device)(f)
grads = so.Gradient.Create(dim=2, device=device)(f) * m
orig_grads = grads.copy()
phi_inv = core.StructuredGrid((256, 256),
spacing=[1.0, 1.0],
device=device)
phi_inv.set_to_identity_lut_()
id = phi_inv.copy()
for i in range(0, 6):
diffuse_x_grad = grads[0].data
diffuse_y_grad = diffuse(grads[1].data, d.data.squeeze())
# Scale back up the gradients
y_scale = (grads[1].max() - grads[1].min()) / (diffuse_y_grad.max() -
diffuse_y_grad.min())
update = core.StructuredGrid.FromGrid(
f,
tensor=torch.stack((diffuse_x_grad, y_scale * diffuse_y_grad), 0),
channels=2)
sample = id.clone() + 20.0 * update
phi_inv = so.ApplyGrid.Create(sample,
pad_mode='border',
device=update.device,
dtype=update.dtype)(phi_inv)
# update the gradients
f = so.ApplyGrid.Create(phi_inv, device=device)(f)
d = so.ApplyGrid.Create(phi_inv, device=device)(d)
m = so.ApplyGrid.Create(phi_inv, device=device)(m)
grads = so.Gradient.Create(dim=2, device=device)(f) * m
print(f'Iter {i}/5 done...')
test = so.ApplyGrid.Create(phi_inv, device=device)(f)
plt.figure()
plt.imshow(d.cpu())
plt.colorbar()
plt.title('Diffusion Coefficient')
plt.figure()
plt.imshow(imgout.cpu())
plt.colorbar()
plt.title('Diffused Gradients')
#
plt.figure()
plt.imshow(grads[1].squeeze().cpu())
plt.colorbar()
plt.title('Starting Gradients')
#
# def_grid = id.copy()
# def_grid.set_to_identity_lut_()
# def_grid += smooth_grads
#
# phi_inv = so.ApplyGrid.Create(def_grid, device=device)(phi_inv)
#
# def_image = so.ApplyGrid.Create(phi_inv, device=device)(f)
#
# plt.figure()
# plt.imshow(def_image.data.squeeze().cpu() - f.data.squeeze().cpu())
# plt.colorbar()
# plt.title('Difference Of Images')
#
# plt.figure()
# plt.imshow(def_image.data.squeeze().cpu())
# plt.title('Deformed Image')
print('something')
def _convert_mat_files(Obj):
rabbitNumber = Obj.rabbitDirectory.split('/')[-1]
matList = sorted(glob.glob(f'{Obj.rabbitDirectory}/externalRecons/AblationImaging/*'))
# Do some string comparison to make sure that this file is a sonication file
matList = [x for x in matList if 'Son' in x]
# Have to keep track of the MIN and MAX of each transfor so we can know the bounding box to resample to
size = []
orgn = []
if not os.path.exists(os.path.join(Obj.rabbitDirectory + '/rawVolumes/TemperatureRecon/RegVols/')):
os.makedirs(os.path.join(Obj.rabbitDirectory + '/rawVolumes/TemperatureRecon/RegVols/'))
for tempFile in matList:
# Create the output folder for all of the timepoint images
outDir = os.path.splitext(tempFile)[0]
outName = outDir.split('/')[-1]
dirOutName = os.path.join(Obj.rabbitDirectory + '/rawVolumes/TemperatureRecon/', outName)
if not os.path.exists(dirOutName):
os.makedirs(dirOutName)
# Have to add mroe logic here because there are varying lengths of numbers FUN
preLength = len(outName.split('_')[0])
outName = outName[:preLength] + '_----' + outName[preLength:]
struct = spi.loadmat(tempFile)
imData = struct['ims']
# Could put some logic here to make sure that they all have a 4th dimension
timePts = np.shape(imData)[-1]
# Get transformation and solve for the affine
# tform = struct['PosDCS']
# print('min: {0} .... '.format(np.min(np.min(np.min(tform, 0), 0), 0).tolist()))
# print('max: {0} .... '.format(np.max(np.max(np.max(tform, 0), 0), 0).tolist()))
# dim = np.shape(tform)[0:3]
# dim = [x - 1 for x in dim]
# print('shape: {0} .... '.format(dim))
# Pull points from the transformation to solve for the affine
# landmarks = []
# landmarks.append([[0, 0, 0], tform[0, 0, 0].tolist()])
# landmarks.append([[dim[0], 0, 0], tform[dim[0], 0, 0].tolist()])
# landmarks.append([[0, dim[1], 0], tform[0, dim[1], 0].tolist()])
# landmarks.append([[dim[0], dim[1], 0], tform[dim[0], dim[1], 0].tolist()])
# landmarks.append([[0, 0, dim[2]], tform[0, 0, dim[2]].tolist()])
# landmarks.append([[dim[0], 0, dim[2]], tform[dim[0], 0, dim[2]].tolist()])
# landmarks.append([[0, dim[1], dim[2]], tform[0, dim[1], dim[2]].tolist()])
# landmarks.append([[dim[0], dim[1], dim[2]], tform[dim[0], dim[1], dim[2]].tolist()])
# Solve for the affine
# affine = apps.SolveAffine(landmarks)
affine = struct['geomInfo'][0,0]['AffineDCS']
temp = core.StructuredGrid(
[imData.shape[0], imData.shape[1], imData.shape[2]],
tensor=torch.tensor(np.real(imData)).permute(-1, 0, 1, 2),
origin= [0, 0, 0],
channels=imData.shape[-1]
)
# temp = ca.Image3D(cc.MakeGrid(np.shape(tform)[0:3], ca.MEM_DEVICE), ca.MEM_DEVICE)
# temp.setOrigin(ca.Vec3Df(0, 0, 0))
# temp.setSpacing(ca.Vec3Df(1, 1, 1))
affGrid = rc.SolveAffineGrid(temp, affine)
size.append(affGrid.size.tolist())
orgn.append(affGrid.origin.tolist())
print('Converting files for {0} time series images .... '.format(timePts), end='')
sys.stdout.flush()
for time in range(0, timePts):
magIm_np = np.real(imData[:, :, :, time])
phsIm_np = np.imag(imData[:, :, :, time])
magIm = core.StructuredGrid(
[magIm_np.shape[0], magIm_np.shape[1], magIm_np.shape[2]],
tensor=torch.tensor(magIm_np).unsqueeze(0).float(),
origin=[0, 0, 0],
channels=1
)
phsIm = core.StructuredGrid(
[phsIm_np.shape[0], phsIm_np.shape[1], phsIm_np.shape[2]],
tensor=torch.tensor(phsIm_np).unsqueeze(0).float(),
origin=[0, 0, 0],
channels=1
)
affFilt = so.AffineTransform.Create(affine=torch.tensor(affine).float())
affMagIm = affFilt(magIm, out_grid=affGrid)
affPhsIm = affFilt(phsIm, out_grid=affGrid)
io.SaveITKFile(affMagIm, dirOutName + '/' + outName + '_real_{0}.nii.gz'.format(str(time).zfill(2)))
io.SaveITKFile(affPhsIm, dirOutName + '/' + outName + '_imag_{0}.nii.gz'.format(str(time).zfill(2)))
if time == 0:
# We need to have the magnitude image for registration
regIm_np = np.abs(imData[:, :, :, time])
regIm = core.StructuredGrid(
[regIm_np.shape[0], regIm_np.shape[1], regIm_np.shape[2]],
tensor=torch.tensor(regIm_np).unsqueeze(0).float(),
origin=[0, 0, 0],
channels=1
)
affRegIm = affFilt(regIm, out_grid=affGrid)
io.SaveITKFile(affRegIm, dirOutName + '/../RegVols/' + outName[0:5] + '.nii.gz')
# print('size: {0} .... '.format(magIm_np.shape), end='')
print('Done')
return size, orgn
def sample_on_histopathology(rabbit, block, img_num, bf_slice):
blockface_dir = f'/hdscratch/ucair/{rabbit}/blockface/{block}/'
histology_dir = f'/hdscratch/ucair/{rabbit}/microscopic/{block}/'
invivo_dir = f'/hdscratch/ucair/{rabbit}/mri/invivo/volumes/deformable/{block}/'
day0_dir = f'/hdscratch/ucair/{rabbit}/mri/day0/volumes/deformable/{block}/'
invivo_mr_out_path = f'{invivo_dir}/IMG_{img_num}/'
day0_mr_out_path = f'{day0_dir}/IMG_{img_num}/'
mic_file = f'{histology_dir}hdf5/{block}_img{img_num}_image.hdf5'
# First need to see if the deformation from histology to blockface exists
# Load the deformation
try:
phi_inv_data = io.LoadITKFile(
f'{histology_dir}deformations/img_{img_num}_deformation_to_blockface.mhd', device=device
)
aff = np.loadtxt(f'{histology_dir}deformations/img_{img_num}_affine_to_blockface.txt')
aff = torch.tensor(aff, device=device, dtype=torch.float32)
except IOError:
raise IOError(f'The full deformation for IMG {img_num} was not found. Please generate and then re-run.')
# Because I can't save 2D deformations at the moment
phi_inv = core.StructuredGrid(
size=phi_inv_data.size[0:2],
spacing=phi_inv_data.spacing[1:3],
origin=phi_inv_data.origin[1:3],
device=phi_inv_data.device,
tensor=phi_inv_data.data.squeeze().permute(2, 0, 1),
channels=2
)
# Apply the inverse affine to the deformation
aff = aff.inverse()
a = aff[0:2, 0:2].float()
t = aff[-0:2, 2].float()
phi_inv.data = torch.matmul(a.unsqueeze(0).unsqueeze(0),
phi_inv.data.permute(list(range(1, 2 + 1)) + [0]).unsqueeze(-1))
phi_inv.data = (phi_inv.data.squeeze() + t).permute([-1] + list(range(0, 2)))
if not os.path.exists(invivo_mr_out_path):
os.makedirs(invivo_mr_out_path)
if not os.path.exists(day0_mr_out_path):
os.makedirs(day0_mr_out_path)
meta_dict = {}
with h5py.File(mic_file, 'r') as f:
mic = f['RawImage/ImageData'][:, ::10, ::10]
for key in f['RawImage'].attrs:
meta_dict[key] = f['RawImage'].attrs[key]
mic = core.StructuredGrid(
mic.shape[1:],
tensor=torch.tensor(mic, dtype=torch.float32, device=device),
spacing=torch.tensor([10.0, 10.0], dtype=torch.float32, device=device),
origin=(torch.tensor(mic.shape[1:]) / 2) * -1,
device=device,
dtype=torch.float32,
channels=3
)
mic = (mic - mic.min()) / (mic.max() - mic.min())
segs = []
segs += [io.LoadITKFile(f'{histology_dir}segmentations/IMG_{img_num}/img_{img_num}_healthy_tissue.nrrd',
device=device)]
if os.path.exists(f'{histology_dir}segmentations/IMG_{img_num}/img_{img_num}_uncertain_region.nrrd'):
segs += [io.LoadITKFile(f'{histology_dir}segmentations/IMG_{img_num}/img_{img_num}_uncertain_region.nrrd',
device=device)]
if os.path.exists(f'{histology_dir}segmentations/IMG_{img_num}/img_{img_num}_ablated_region.nrrd'):
segs += [io.LoadITKFile(f'{histology_dir}segmentations/IMG_{img_num}/img_{img_num}_ablated_region.nrrd',
device=device)]
# Apply the deformation to the microscopic image
deformed_histology = so.ApplyGrid(phi_inv, device=device)(mic, phi_inv)
for i, vol in enumerate(segs):
segs[i] = so.ApplyGrid(phi_inv, device=device)(vol, phi_inv)
# Load the blockface
blockface = io.LoadITKFile(f'{blockface_dir}volumes/raw/difference_volume.mhd', device=device)
sd = 'cuda:0'
mr_invivo_t1_slice = get_mr_slice(f'{invivo_dir}/invivo_ce_t1_to_{block}.mhd', blockface, bf_slice - 1, sd)
# mr_invivo_t2_slice = get_mr_slice(f'{invivo_dir}/invivo_t2_to_{block}.mhd', blockface, bf_slice - 1, sd)
# mr_invivo_adc_slice = get_mr_slice(f'{invivo_dir}/invivo_adc_to_{block}.mhd', blockface, bf_slice - 1, sd)
mr_invivo_npv_slice = get_mr_slice(f'{invivo_dir}/invivo_npv_to_{block}.mhd', blockface, bf_slice - 1, sd)
mr_d0_npv_slice = get_mr_slice(f'/hdscratch/ucair/18_062/mri/day0/volumes/deformable/block04/day0_npv_to_{block}.mhd', blockface, bf_slice - 1, sd)
mr_d0_t1_slice = get_mr_slice(
f'/hdscratch/ucair/18_062/mri/day0/volumes/deformable/block04/day0_t1_to_{block}.mhd', blockface, bf_slice - 1,
sd)
mr_d0_t1_nc_slice = get_mr_slice(
f'/hdscratch/ucair/18_062/mri/day0/volumes/deformable/block04/day0_t1_nc_to_{block}.mhd', blockface, bf_slice - 1,
sd)
mr_log_ctd_slice = get_mr_slice(
f'/hdscratch/ucair/18_062/mri/day0/volumes/deformable/block04/day0_log_ctd_to_{block}.mhd', blockface, bf_slice - 1,
sd)
mr_log_ctd_slice.data = torch.exp(mr_log_ctd_slice.data)
def create_circular_mask(h, w, center=None, radius=None):
if center is None: # use the middle of the image
center = (int(w / 2), int(h / 2))
if radius is None: # use the smallest distance between the center and image walls
radius = min(center[0], center[1], w - center[0], h - center[1])
Y, X = np.ogrid[:h, :w]
dist_from_center = np.sqrt((X - center[0]) ** 2 + (Y - center[1]) ** 2)
mask = dist_from_center <= radius
return mask
circle_mask = create_circular_mask(2586, 3426, center=[1831, 1100], radius=700)
mr_log_ctd_slice.data = mr_log_ctd_slice.data * torch.tensor(circle_mask, device=device).float()
# mr_day0_t2_slice = get_mr_slice(f'{day0_dir}/day0_t2_to_{block}.mhd', blockface, bf_slice - 1, sd)
# mr_day0_ctd_slice = get_mr_slice(f'{day0_dir}/day0_ctd_to_{block}.mhd', blockface, bf_slice - 1, sd)
# mr_day0_t1_slice = get_mr_slice(f'{day0_dir}/day0_ce_t1_to_{block}.mhd', blockface, bf_slice - 1, sd)
# mr_day0_npv_slice = get_mr_slice(f'{day0_dir}/day0_npv_to_{block}.mhd', blockface, bf_slice - 1, sd)
# io.SaveITKFile(mr_invivo_t1_slice, f'{invivo_mr_out_path}/invivo_ce_t1_as_bf_img_{img_num}.mhd')
# io.SaveITKFile(mr_invivo_t2_slice, f'{invivo_mr_out_path}/invivo_t2_as_bf_img_{img_num}.mhd')
# io.SaveITKFile(mr_invivo_adc_slice, f'{invivo_mr_out_path}/invivo_adc_as_bf_img_{img_num}.mhd')
# io.SaveITKFile(mr_invivo_npv_slice, f'{invivo_mr_out_path}/invivo_npv_as_bf_img_{img_num}.mhd')
# io.SaveITKFile(mr_day0_t2_slice, f'{day0_mr_out_path}/day0_t2_as_bf_img_{img_num}.mhd')
# io.SaveITKFile(mr_day0_ctd_slice, f'{day0_mr_out_path}/day0_ctd_as_bf_img_{img_num}.mhd')
# io.SaveITKFile(mr_day0_t1_slice, f'{day0_mr_out_path}/day0_ce_t1_as_bf_img_{img_num}.mhd')
# io.SaveITKFile(mr_day0_npv_slice, f'{day0_mr_out_path}/day0_ce_t1_as_bf_img_{img_num}.mhd')
del blockface, phi_inv, phi_inv_data
torch.cuda.empty_cache()
histology_seg = core.StructuredGrid.FromGrid(segs[0], channels=1)
for seg in segs:
histology_seg += seg
ctd_slice = mr_log_ctd_slice.data[0].cpu()
ctd_mask = np.logical_and(ctd_slice > 10.0, circle_mask == 1)
masked_thermal = np.ma.masked_where(ctd_mask == 0.0, ctd_slice.numpy()).squeeze()
plt.figure()
# plt.imshow(t1_nc.data.cpu()[0, slice_n].squeeze(), aspect=1.0 / aspect, cmap='gray')
plt.imshow(masked_thermal, cmap='jet', vmin=10, vmax=240)
plt.gca().patch.set_facecolor([0, 0, 0, 0])
plt.axis('off')
zero_slice = np.zeros_like(ctd_slice)
masked_slice = np.ma.masked_where(zero_slice == 0.0, zero_slice).squeeze()
plt.figure()
plt.imshow(masked_slice.squeeze(), cmap='gray')
npv_contours = measure.find_contours(ctd_slice.data.squeeze().cpu().numpy() * circle_mask, 240)
for contour in npv_contours:
plt.plot(contour[:, 1], contour[:, 0], color='red', linewidth=1.8)
plt.axis('off')
plt.gca().patch.set_facecolor([0, 0, 0, 0])
plt.savefig(f'{invivo_mr_out_path}/day0_ctd_contour.png', dpi=600, bbox_inches='tight', pad_inches=0, transparent=True)
save = True
# Generate the figures without masking
generate_figures(
mr_d0_t1_slice, segs, out_path=invivo_mr_out_path, base_name='day0_ce_t1_no_mask', save=save
)
generate_figures(
mr_invivo_t1_slice, segs, out_path=invivo_mr_out_path, base_name='deformed_invivo_ce_t1_MR_no_mask', save=save
)
histology_image = histology_seg * deformed_histology
mr_invivo_t1_slice = histology_seg * mr_invivo_t1_slice
mr_d0_t1_slice = histology_seg * mr_d0_t1_slice
# mr_invivo_t2_slice = histology_seg * mr_invivo_t2_slice
# mr_day0_t2_slice = histology_seg * mr_day0_t2_slice
# mr_day0_ctd_slice = histology_seg * mr_day0_ctd_slice
# mr_day0_t1_slice = histology_seg * mr_day0_t1_slice
# mr_day0_npv_slice = histology_seg * mr_day0_npv_slice
# mr_day0_ctd_slice.data[mr_day0_ctd_slice.data < 0.5] = 0.0
# mr_day0_ctd_slice.data[mr_day0_ctd_slice.data >= 0.5] = 1.0
#
# mr_day0_npv_slice.data[mr_day0_npv_slice.data < 0.5] = 0.0
# mr_day0_npv_slice.data[mr_day0_npv_slice.data >= 0.5] = 1.0
hist_map = core.StructuredGrid.FromGrid(segs[0])
for i, seg in enumerate(segs, 1):
hist_map = hist_map + (i * seg)
if len(segs) == 3:
color_map = ListedColormap(['k', 'crimson', 'lime', 'gold'])
plt.figure()
plt.imshow(hist_map.data.squeeze().cpu(), cmap=color_map)
elif len(segs) == 2:
color_map = ListedColormap(['k', 'crimson', 'lime'])
plt.figure()
plt.imshow(hist_map.data.squeeze().cpu(), cmap=color_map)
else:
color_map = ListedColormap(['k', 'crimson'])
plt.figure()
plt.imshow(hist_map.data.squeeze().cpu(), cmap=color_map)
plt.axis('off')
plt.show()
plt.pause(1.0)
if save:
plt.savefig(f'{invivo_mr_out_path}/histology_segmentation_map.png', dpi=600, bbox_inches='tight', pad_inches=0)
plt.savefig(f'{day0_mr_out_path}/histology_segmentation_map.png', dpi=600, bbox_inches='tight', pad_inches=0)
plt.close('all')
# Generate the figure for showing the NPV on the Day0 MRI CE T1
# contours = measure.find_contours(mr_day0_npv_slice.data.squeeze().cpu().numpy(), 0.5)
# Plot the image with contours
# plt.figure()
# plt.imshow(mr_day0_t1_slice.data.permute(1, 2, 0).squeeze().cpu(), cmap='gray')
# for contour in contours:
# plt.plot(contour[:, 1], contour[:, 0], color='blue', linewidth=0.8)
# plt.axis('off')
# plt.show()
# plt.pause(1.0)
# if save:
# plt.savefig(f'{invivo_mr_out_path}/day0_npv_with_contours.png', dpi=600, bbox_inches='tight', pad_inches=0)
#
# plt.close('all')
# generate_figures(
# mr_day0_npv_slice, segs, out_path=day0_mr_out_path, base_name='deformed_npv', save=save,
# extra_cont=contours
# )
generate_figures(
histology_image, segs, out_path=invivo_mr_out_path, base_name='deformed_histology', save=save
)
generate_figures(
mr_invivo_t1_slice, segs, out_path=invivo_mr_out_path, base_name='deformed_invivo_ce_t1_MR', save=save
)
generate_figures(
mr_d0_t1_slice, segs, out_path=invivo_mr_out_path, base_name='day0_ce_t1', save=save
)
# generate_figures(
# mr_invivo_t2_slice, segs, out_path=invivo_mr_out_path, base_name='deformed_invivo_t2_MR', save=save
# )
# generate_figures(
# mr_invivo_adc_slice, segs, out_path=invivo_mr_out_path, base_name='deformed_invivo_adc_MR', save=save
# )
# generate_figures(
# mr_day0_t2_slice, segs, out_path=day0_mr_out_path, base_name='deformed_day0_t2_MR', save=save
# )
# generate_figures(
# mr_day0_ctd_slice, segs, out_path=day0_mr_out_path, base_name='deformed_day0_ctd_MR', save=save
# )
#
# generate_figures(
# mr_day0_t1_slice, segs, out_path=day0_mr_out_path, base_name='deformed_day0_ce_t1_MR', save=True
# )
npv_contours = measure.find_contours(mr_invivo_npv_slice.data.squeeze().cpu().numpy(), 0.5)
zero_im = np.zeros_like(mr_invivo_npv_slice.data.squeeze().cpu().numpy())
fig = plt.figure()
# plt.imshow(mr_invivo_t1_slice.data.cpu().squeeze())
ax = fig.add_subplot()
ax.set_xlim([0, zero_im.shape[1]])
ax.set_ylim([0, zero_im.shape[0]])
for contour in npv_contours:
plt.plot(contour[:, 1], contour[:, 0], color=matplotlib._cm._tab10_data[2], linewidth=3.0)
ax.invert_yaxis()
ax.set_aspect(1)
plt.axis('off')
plt.savefig(f'{invivo_mr_out_path}/NPV_contours.png', dpi=600, bbox_inches='tight', pad_inches=0, transparent=True)
npv_contours = measure.find_contours(mr_d0_npv_slice.data.squeeze().cpu().numpy(), 0.5)
zero_im = np.zeros_like(mr_d0_npv_slice.data.squeeze().cpu().numpy())
fig = plt.figure()
# plt.imshow(mr_invivo_t1_slice.data.cpu().squeeze())
ax = fig.add_subplot()
ax.set_xlim([0, zero_im.shape[1]])
ax.set_ylim([0, zero_im.shape[0]])
for contour in npv_contours:
plt.plot(contour[:, 1], contour[:, 0], color=matplotlib._cm._tab10_data[0], linewidth=3.0)
ax.invert_yaxis()
ax.set_aspect(1)
plt.axis('off')
plt.savefig(f'{invivo_mr_out_path}/D0_NPV_contours.png', dpi=600, bbox_inches='tight', pad_inches=0, transparent=True)
if len(segs) == 3:
hst_contours = measure.find_contours(segs[-1].data.squeeze().cpu().numpy(), 0.5)
fig = plt.figure()
ax = fig.add_subplot()
ax.set_xlim([0, zero_im.shape[1]])
ax.set_ylim([0, zero_im.shape[0]])
for contour in hst_contours:
plt.plot(contour[:, 1], contour[:, 0], color=matplotlib._cm._tab10_data[6], linewidth=3.0)
ax.invert_yaxis()
ax.set_aspect(1)
plt.axis('off')
plt.savefig(f'{invivo_mr_out_path}/HST_contours.png', dpi=600, bbox_inches='tight', pad_inches=0, transparent=True)
plt.close('all')
print('Done')
def generate_image_volume(rabbit, block, base_dir='/hdscratch/ucair/'):
blockface_dir = f'{base_dir}{rabbit}/blockface/{block}/'
histology_dir = f'{base_dir}{rabbit}/microscopic/{block}/'
def_dir = f'{histology_dir}deformations/'
out_dir = f'{histology_dir}/volume/raw/'
if not os.path.exists(out_dir):
os.makedirs(out_dir)
raw_images = glob.glob(f'{histology_dir}/raw/*_image.tif')
if not raw_images:
raw_images += sorted(raw_images +
glob.glob(f'{histology_dir}/raw/*_image.jpg'))
image_nums = sorted(
[int(x.split('/')[-1].split('_')[1]) for x in raw_images])
deformed_images = []
for im in image_nums:
print(f'Loading and deforming image {im} ... ', end='')
# Load the healthy segmentation as a reference
healthy = io.LoadITKFile(
f'{histology_dir}segmentations/IMG_{im:03d}/img_{im:03d}_healthy_tissue.nrrd',
device=device)
try:
ablation = io.LoadITKFile(
f'{histology_dir}segmentations/IMG_{im:03d}/img_{im:03d}_ablated_region.nrrd',
device=device)
except RuntimeError:
ablation = (healthy * 0.0).copy()
try:
transition = io.LoadITKFile(
f'{histology_dir}segmentations/IMG_{im:03d}/img_{im:03d}_uncertain_region.nrrd',
device=device)
except RuntimeError:
transition = (healthy * 0.0).copy()
combined = ablation + transition + healthy
combined.data = (combined.data >= 0.5).float()
# Load the affine
aff = np.loadtxt(
glob.glob(f'{def_dir}img_{im:03d}_affine_to_blockface.txt')[0])
aff = torch.tensor(aff, device=device, dtype=torch.float32)
# Load the deformation
deformation_data = io.LoadITKFile(
f'{def_dir}/img_{im:03d}_deformation_to_blockface.mhd',
device=device)
deformation = core.StructuredGrid(
size=deformation_data.size[0:2],
spacing=deformation_data.spacing[1:3],
origin=deformation_data.origin[1:3],
device=deformation_data.device,
tensor=deformation_data.data.squeeze().permute(2, 0, 1),
channels=2)
# Apply the inverse affine to the grid
aff = aff.inverse()
a = aff[0:2, 0:2].float()
t = aff[-0:2, 2].float()
deformation.data = torch.matmul(
a.unsqueeze(0).unsqueeze(0),
deformation.data.permute(list(range(1, 2 + 1)) +
[0]).unsqueeze(-1))
deformation.data = (deformation.data.squeeze() +
t).permute([-1] + list(range(0, 2)))
# Load the actual image
#### Apply the affine to the image
mic_file = f'{histology_dir}hdf5/{block}_img{im:03d}_image.hdf5'
meta_dict = {}
with h5py.File(mic_file, 'r') as f:
mic = f['RawImage/ImageData'][:, ::10, ::10]
for key in f['RawImage'].attrs:
meta_dict[key] = f['RawImage'].attrs[key]
mic = core.StructuredGrid(mic.shape[1:],
tensor=torch.tensor(mic,
dtype=torch.float32,
device=device),
spacing=torch.tensor([10.0, 10.0],
dtype=torch.float32,
device=device),
device=device,
dtype=torch.float32,
channels=3)
mic = (mic - mic.min()) / (mic.max() - mic.min())
mic.data = mic.data * combined.data
def_mic = so.ApplyGrid.Create(deformation, device=device)(mic,
deformation)
# Need to put the mic image in the right slice
z_or = (im - 1) * 0.05
def_mic = core.StructuredGrid(
[2] + list(def_mic.shape()[1:]),
tensor=def_mic.data.unsqueeze(1).repeat(1, 2, 1, 1),
spacing=torch.tensor(
[0.05, def_mic.spacing[0], def_mic.spacing[1]],
dtype=torch.float32,
device=device),
origin=torch.tensor([z_or, def_mic.origin[0], def_mic.origin[1]],
dtype=torch.float32,
device=device),
device=device,
dtype=torch.float32,
channels=3)
deformed_images.append(def_mic.copy())
del def_mic, mic, deformation
torch.cuda.empty_cache()
print('done')
# Now need to load the blockface volume
block_vol = io.LoadITKFile(
f'{blockface_dir}volumes/raw/difference_volume.mhd', device='cpu')
# image_vol = core.StructuredGrid(
# size=block_vol.size,
# spacing=block_vol.spacing,
# origin=block_vol.origin,
# device=block_vol.device,
# channels=3
# )
image_z_spacing = np.diff(np.array(image_nums)).mean() * 0.05
image_z_origin = image_nums[0] * 0.05
image_vol = core.StructuredGrid(
size=[len(image_nums)] + list(block_vol.size[-2:].numpy()),
spacing=[image_z_spacing] + list(block_vol.spacing[-2:].numpy()),
origin=[image_z_origin] + list(block_vol.origin[-2:].numpy()),
device=block_vol.device,
channels=3)
for ii, im in enumerate(image_nums):
image_vol.data[:, ii] = deformed_images[ii].data[:, 0].clone().cpu()
io.SaveITKFile(image_vol, f'{out_dir}/{block}_image_volume.mhd')
del image_vol, deformed_images, block_vol
torch.cuda.empty_cache()
def generate_segmentation_volume(rabbit, block, base_dir='/hdscratch/ucair/'):
blockface_dir = f'{base_dir}{rabbit}/blockface/{block}/'
histology_dir = f'{base_dir}{rabbit}/microscopic/{block}/'
def_dir = f'{histology_dir}deformations/'
out_dir = f'{histology_dir}/volume/raw/'
if not os.path.exists(out_dir):
os.makedirs(out_dir)
raw_images = glob.glob(f'{histology_dir}/raw/*_image.tif')
if not raw_images:
raw_images += sorted(raw_images +
glob.glob(f'{histology_dir}/raw/*_image.jpg'))
image_nums = sorted(
[int(x.split('/')[-1].split('_')[1]) for x in raw_images])
deformed_ablation_segs = []
deformed_combined_segs = []
deformed_images = []
for im in image_nums:
print(f'Loading and deforming image {im} ... ', end='')
# Load the healthy segmentation as a reference
healthy = io.LoadITKFile(
f'{histology_dir}segmentations/IMG_{im:03d}/img_{im:03d}_healthy_tissue.nrrd',
device=device)
# Load the affine
aff = np.loadtxt(
glob.glob(f'{def_dir}img_{im:03d}_affine_to_blockface.txt')[0])
aff = torch.tensor(aff, device=device, dtype=torch.float32)
# Load the deformation
deformation_data = io.LoadITKFile(
f'{def_dir}/img_{im:03d}_deformation_to_blockface.mhd',
device=device)
deformation = core.StructuredGrid(
size=deformation_data.size[0:2],
spacing=deformation_data.spacing[1:3],
origin=deformation_data.origin[1:3],
device=deformation_data.device,
tensor=deformation_data.data.squeeze().permute(2, 0, 1),
channels=2)
# Apply the inverse affine to the grid
aff = aff.inverse()
a = aff[0:2, 0:2].float()
t = aff[-0:2, 2].float()
deformation.data = torch.matmul(
a.unsqueeze(0).unsqueeze(0),
deformation.data.permute(list(range(1, 2 + 1)) +
[0]).unsqueeze(-1))
deformation.data = (deformation.data.squeeze() +
t).permute([-1] + list(range(0, 2)))
try:
ablation = io.LoadITKFile(
f'{histology_dir}segmentations/IMG_{im:03d}/img_{im:03d}_ablated_region.nrrd',
device=device)
except RuntimeError:
ablation = (healthy * 0.0).copy()
try:
transition = io.LoadITKFile(
f'{histology_dir}segmentations/IMG_{im:03d}/img_{im:03d}_uncertain_region.nrrd',
device=device)
except RuntimeError:
transition = (healthy * 0.0).copy()
# Load the actual image
#### Apply the affine to the image
mic_file = f'{histology_dir}hdf5/{block}_img{im:03d}_image.hdf5'
meta_dict = {}
with h5py.File(mic_file, 'r') as f:
mic = f['RawImage/ImageData'][:, ::10, ::10]
for key in f['RawImage'].attrs:
meta_dict[key] = f['RawImage'].attrs[key]
mic = core.StructuredGrid(mic.shape[1:],
tensor=torch.tensor(mic,
dtype=torch.float32,
device=device),
spacing=torch.tensor([10.0, 10.0],
dtype=torch.float32,
device=device),
device=device,
dtype=torch.float32,
channels=3)
mic = (mic - mic.min()) / (mic.max() - mic.min())
if not os.path.exists(
f'{histology_dir}deformable/img_{im:03d}_to_blockface/'):
os.makedirs(
f'{histology_dir}deformable/img_{im:03d}_to_blockface/')
def_mic = so.ApplyGrid.Create(deformation, device=device)(mic,
deformation)
# Need to put the mic image in the right slice
z_or = (im - 1) * 0.05
def_mic = core.StructuredGrid(
[2] + list(def_mic.shape()[1:]),
tensor=def_mic.data.unsqueeze(1).repeat(1, 2, 1, 1),
spacing=torch.tensor(
[0.05, def_mic.spacing[0], def_mic.spacing[1]],
dtype=torch.float32,
device=device),
origin=torch.tensor([z_or, def_mic.origin[0], def_mic.origin[1]],
dtype=torch.float32,
device=device),
device=device,
dtype=torch.float32,
channels=3)
io.SaveITKFile(
def_mic,
f'{histology_dir}deformable/img_{im:03d}_to_blockface/img_{im:03d}_to_blockface.mhd'
)
combined = ablation + transition
combined.data = (combined.data >= 0.5).float()
def_ablation = so.ApplyGrid.Create(deformation,
device=device)(ablation,
deformation)
def_combined = so.ApplyGrid.Create(deformation,
device=device)(combined,
deformation)
deformed_ablation_segs.append(def_ablation.copy())
deformed_combined_segs.append(def_combined.copy())
deformed_images.append(def_mic.copy())
del def_ablation, def_combined, def_mic, combined, mic, deformation
torch.cuda.empty_cache()
print('done')
# Now need to load the blockface volume
block_vol = io.LoadITKFile(
f'{blockface_dir}volumes/raw/difference_volume.mhd', device='cpu')
image_vol = core.StructuredGrid(size=block_vol.size,
spacing=block_vol.spacing,
origin=block_vol.origin,
device=block_vol.device,
channels=3)
for ii, im in enumerate(image_nums):
image_vol.data[:, im - 1] = deformed_images[ii].data[:,
0].clone().cpu()
io.SaveITKFile(image_vol, f'{out_dir}/{block}_image_volume.mhd')
del image_vol, deformed_images
torch.cuda.empty_cache()
single_vol = core.StructuredGrid(size=block_vol.size,
spacing=block_vol.spacing,
origin=block_vol.origin,
device=block_vol.device,
channels=1)
for ii, im in enumerate(image_nums):
single_vol.data[:, im -
1] = deformed_ablation_segs[ii].data[:].clone().cpu()
io.SaveITKFile(single_vol,
f'{out_dir}/{block}_ablation_segmentation_no_interp.mhd')
del single_vol
torch.cuda.empty_cache()
# Now need to load the blockface volume
ablation_vol = core.StructuredGrid(size=block_vol.size,
spacing=block_vol.spacing,
origin=block_vol.origin,
device=device,
channels=1)
for ii, im in enumerate(image_nums):
if ii == len(image_nums) - 1:
continue
next_slice = image_nums[ii + 1]
for s, slice in enumerate(range(im, next_slice + 1)):
step = 1.0 / (next_slice - im)
cur_alpha = 1.0 - (s * step)
next_alpa = 0.0 + (s * step)
ablation_vol.data[:, slice] = (
deformed_ablation_segs[ii].data *
cur_alpha) + (deformed_ablation_segs[ii + 1].data * next_alpa)
gauus_ablation = so.Gaussian.Create(1,
10, [0.1, 2, 2],
dim=3,
device=device)(ablation_vol)
io.SaveITKFile(gauus_ablation,
f'{out_dir}/{block}_ablation_segmentation.mhd')
del gauus_ablation, deformed_ablation_segs
torch.cuda.empty_cache()
combined_vol = core.StructuredGrid(size=block_vol.size,
spacing=block_vol.spacing,
origin=block_vol.origin,
device=block_vol.device,
channels=1)
for ii, im in enumerate(image_nums):
if ii == len(image_nums) - 1:
continue
next_slice = image_nums[ii + 1]
for s, slice in enumerate(range(im, next_slice + 1)):
step = 1.0 / (next_slice - im)
cur_alpha = 1.0 - (s * step)
next_alpa = 0.0 + (s * step)
combined_vol.data[:, slice] = (
deformed_combined_segs[ii].data *
cur_alpha) + (deformed_combined_segs[ii + 1].data * next_alpa)
# gauus_ablation = so.Gaussian.Create(1, 10, [0.1, 2, 2], dim=3, device=device)(ablation_vol)
io.SaveITKFile(
combined_vol,
f'{out_dir}/{block}_ablation_and_transition_segmentation.mhd')
del block_vol
del combined_vol
torch.cuda.empty_cache()
def process_mic(micicroscopic, mic_seg_file, blockface, label, device='cpu'):
meta_dict = {}
try:
with h5py.File(micicroscopic, 'r') as f:
mic = f['ImageData'][1, ::10, ::10]
except KeyError:
with h5py.File(micicroscopic, 'r') as f:
mic = f['RawImage/ImageData'][1, ::10, ::10]
for key in f['RawImage'].attrs:
meta_dict[key] = f['RawImage'].attrs[key]
with h5py.File(mic_seg_file, 'r') as f:
mic_seg = f['ImageData'][0, ::10, ::10]
mic = core.StructuredGrid(
mic.shape,
tensor=torch.tensor(mic, dtype=torch.float32, device=device).unsqueeze(0),
spacing=torch.tensor([10.0, 10.0], dtype=torch.float32, device=device),
device=device,
dtype=torch.float32,
channels=1
)
mic_seg = core.StructuredGrid(
mic_seg.shape,
tensor=torch.tensor(mic_seg, dtype=torch.float32, device=device).unsqueeze(0),
spacing=torch.tensor([10.0, 10.0], dtype=torch.float32, device=device),
device=device,
dtype=torch.float32,
channels=1
)
# mic_seg.data = (mic_seg.data <= 0.5).float()
if 'Affine' in meta_dict:
opt_affine = torch.tensor(meta_dict['Affine'], dtype=torch.float32, device=device)
optaff_filter = AffineTransform.Create(affine=opt_affine, device=device)
aff_mic_image = optaff_filter(mic, blockface)
aff_mic_label = optaff_filter(mic_seg, blockface)
return aff_mic_image, aff_mic_label, opt_affine
points = torch.tensor(
LandmarkPicker([mic[0].cpu(), blockface[1].cpu()]),
dtype=torch.float32,
device=device
).permute(1, 0, 2)
# Change to real coordinates
points *= torch.cat([mic.spacing[None, None, :], blockface.spacing[None, None, :]], 0)
points += torch.cat([mic.origin[None, None, :], blockface.origin[None, None, :]], 0)
aff_filter = AffineTransform.Create(points[1], points[0], device=device)
affine = torch.eye(3, device=device, dtype=torch.float32)
affine[0:2, 0:2] = aff_filter.affine
affine[0:2, 2] = aff_filter.translation
# aff_mic_image = aff_filter(mic, blockface)
aff_mic_seg = aff_filter(mic_seg, blockface)
# Do some additional registration just to make sure it is in the right spot
similarity = so.L2Similarity.Create(device=device)
model = so.AffineIntensity.Create(similarity, device=device)
# Create the optimizer
optimizer = optim.SGD([
{'params': model.affine, 'lr': 1.0e-11},
{'params': model.translation, 'lr': 1.0e-12}], momentum=0.9, nesterov=True
)
energy = []
for epoch in range(0, 1000):
optimizer.zero_grad()
loss = model(
label.data, aff_mic_seg.data
)
energy.append(loss.item())
print(f'===> Iteration {epoch:3} Energy: {loss.item():.3f}')
loss.backward() # Compute the gradients
optimizer.step() #
# if epoch >= 2:
if epoch > 10 and np.mean(energy[-10:]) - energy[-1] < 0.01:
break
itr_affine = torch.eye(3, device=device, dtype=torch.float32)
itr_affine[0:2, 0:2] = model.affine
itr_affine[0:2, 2] = model.translation
opt_affine = torch.matmul(itr_affine.detach(), affine)
# Create a new resample filter to make sure everything works
optaff_filter = AffineTransform.Create(affine=opt_affine, device=device)
aff_mic_image = optaff_filter(mic, blockface)
aff_mic_label = optaff_filter(mic_seg, blockface)
return aff_mic_image, aff_mic_label, opt_affine
def MultiscaleElast(regObj, opt):
device = opt.device
try:
src_name = regObj.I0
tar_name = regObj.I1
regObj.I1 = io.LoadITKFile(regObj.I1, device)
regObj.I0 = io.LoadITKFile(regObj.I0, device)
except RuntimeError:
src_name = _getFilePath('Source Image File (ElastReg)',
initialdir='/home/sci/blakez/ucair/')
tar_name = _getFilePath('Target Image File (ElastReg)',
initialdir='/home/sci/blakez/ucair/')
regObj.I1 = io.LoadITKFile(tar_name, device)
regObj.I0 = io.LoadITKFile(src_name, device)
# Check if the croppedImage exists in the object - temperature object wont have it
if regObj.croppedImage != 'None':
# Load the cropped image which contains the region of interest
try:
crp_im = io.LoadITKFile(regObj.croppedImage, device)
except RuntimeError:
regObj.croppedImage = _getFilePath(
'Cropped Image File', initialdir='/home/sci/blakez/ucair/')
crp_im = io.LoadITKFile(regObj.croppedImage, device)
# Resample the images onto the ROI grid
regObj.I1 = so.ResampleWorld.Create(crp_im, device=device)(regObj.I1)
regObj.I0 = so.ResampleWorld.Create(crp_im, device=device)(regObj.I0)
del crp_im
if hasattr(regObj, 'unionSize'):
# Make the grid for the union and resample the two images onto that grid
# unionGrid = cc.MakeGrid(regObj.unionSize, regObj.unionSpacing, regObj.unionOrigin)
unionIm = core.StructuredGrid(size=regObj.unionSize,
origin=regObj.unionOrigin,
spacing=regObj.unionSpacing,
channels=1)
regObj.I0 = so.ResampleWorld.Create(unionIm, device=device)(regObj.I0)
regObj.I1 = so.ResampleWorld.Create(unionIm, device=device)(regObj.I1)
# Make sure that the registration is only happening where both of them are not 0
regObj.I0.data[regObj.I1.data == 0.0] = 0.0
regObj.I1.data[regObj.I0.data == 0.0] = 0.0
# Make sure that both images are from 0 to 1
regObj.I0 = regObj.I0 / regObj.I0.max()
regObj.I1 = regObj.I1 / regObj.I1.max()
# Median filter if the regObj has a kernel size for it
if regObj.medianFilterSize != 'None':
regObj.I0 = rc.MedianFilter_I(regObj.I0, regObj.medianFilterSize)
regObj.I1 = rc.MedianFilter_I(regObj.I1, regObj.medianFilterSize)
if regObj.tvWeight != 'None':
regObj.I0 = rc.TVFilter_I(regObj.I0, regObj.tvWeight)
regObj.I1 = rc.TVFilter_I(regObj.I1, regObj.tvWeight)
# Variance Equalize volumes
if regObj.sourceVarESig != 'None':
kernel_size = regObj.sourceVarESig * 3
regObj.I0 = so.VarianceEqualize.Create(kernel_size,
regObj.sourceVarESig,
device=device)(regObj.I0)
if regObj.targetVarESig != 'None':
kernel_size = regObj.sourceVarESig * 3
regObj.I1 = so.VarianceEqualize.Create(kernel_size,
regObj.sourceVarESig,
device=device)(regObj.I1)
# Check for masks: assume they mask out the water or bladder
if regObj.I1_mask != 'None':
try:
mask = io.LoadITKFile(regObj.I1_mask, device)
except RuntimeError:
regObj.I1_mask = _getFilePath('Mask Image File (Target)',
initialdir='/')
mask = io.LoadITKFile(regObj.I1_mask, device)
mask = so.ResampleWorld.Create(regObj.I1, device=opt.device)(mask)
regObj.I1 = regObj.I1 * ((1 - mask) * -1)
del mask
torch.cuda.empty_cache()
if regObj.I0_mask != 'None':
try:
mask = io.LoadITKFile(regObj.I0_mask, device)
except RuntimeError:
regObj.I0_mask = _getFilePath('Mask Image File (Source)',
initialdir='/')
mask = io.LoadITKFile(regObj.I0_mask, device)
mask = so.ResampleWorld.Create(regObj.I0, device=opt.device)(mask)
regObj.I0 = regObj.I0 * ((1 - mask) * -1)
del mask
torch.cuda.empty_cache()
# Gaussian blur the images - blue after masking so the edges are not as sharp
if regObj.gaussSigma != 'None':
blur = so.Gaussian.Create(1,
regObj.gaussKernel * 3,
regObj.gaussSigma * 3,
dim=3,
device=device,
dtype=torch.float32)
regObj.I0 = blur(regObj.I0)
regObj.I1 = blur(regObj.I1)
deformation = regObj.I1.clone()
deformation.set_to_identity_lut_()
deformation_list = []
# Create a grid composer
composer = so.ComposeGrids(device=device,
dtype=torch.float32,
padding_mode='border')
#
for i, s in enumerate(regObj.scales):
scale_source = regObj.I0.set_size(regObj.I1.size // s, inplace=False)
scale_target = regObj.I1.set_size(regObj.I1.size // s, inplace=False)
deformation = deformation.set_size(regObj.I1.size // s, inplace=False)
# Apply the deformation to the source image
scale_source = so.ApplyGrid(deformation)(scale_source)
# Create the matching term
similarity = so.L2Similarity.Create(dim=3, device=device)
# Create the smoothing operator
operator = so.FluidKernel.Create(
scale_target,
device=device,
alpha=1.0,
beta=0.0,
gamma=0.001,
)
# Create the regularizer
regularizer = so.NormGradient.Create(weight=regObj.regWeight[i],
device=device,
dtype=regObj.I1.dtype,
dim=3)
# Now register the source and the gad volume
interday = st.IterativeMatch.Create(
source=scale_source,
target=scale_target,
similarity=similarity,
operator=operator,
device=device,
step_size=regObj.stepSize[i],
regularization=regularizer,
incompressible=regObj.incompressible)
energy = [interday.initial_energy]
print(f'Iteration: 0 Energy: {interday.initial_energy}')
for i in range(1, regObj.Niter[i]):
energy.append(interday.step())
print(f'Iteration: {i} Energy: {energy[-1]}')
# if energy[-1] > energy[-2] - (3e-4 * energy[-2]):
# break
deformation = interday.get_field()
deformation_list.append(deformation.clone().set_size(regObj.I1.size,
inplace=False))
deformation = composer(deformation_list[::-1])
regObj.I0 = src_name
regObj.I1 = tar_name
return deformation
def register_histopathology_to_blockface(rabbit, block, img_num, bf_slice):
blockface_dir = f'/hdscratch/ucair/{rabbit}/blockface/{block}/'
histology_dir = f'/hdscratch/ucair/{rabbit}/microscopic/{block}/'
out_dir = f'{histology_dir}deformations/'
if not os.path.exists(out_dir):
os.makedirs(out_dir)
# if os.path.exists(f'{out_dir}/img_{img_num}_deformation_to_blockface.mhd'):
# return
# Load and make the histopathology segmentation
segs = []
segs += [
io.LoadITKFile(
f'{histology_dir}segmentations/IMG_{img_num}/img_{img_num}_healthy_tissue.nrrd',
device=device)
]
if os.path.exists(
f'{histology_dir}segmentations/IMG_{img_num}/img_{img_num}_ablated_region.nrrd'
):
segs += [
io.LoadITKFile(
f'{histology_dir}segmentations/IMG_{img_num}/img_{img_num}_ablated_region.nrrd',
device=device)
]
if os.path.exists(
f'{histology_dir}segmentations/IMG_{img_num}/img_{img_num}_uncertain_region.nrrd'
):
segs += [
io.LoadITKFile(
f'{histology_dir}segmentations/IMG_{img_num}/img_{img_num}_uncertain_region.nrrd',
device=device)
]
histology_seg = core.StructuredGrid.FromGrid(segs[0], channels=1)
for seg in segs:
histology_seg += seg
try:
blockface_seg = io.LoadITKFile(
f'{blockface_dir}volumes/raw/hd_labels/{block}_hdlabel_volume.nrrd',
device=device)
except:
blockface_seg = io.LoadITKFile(
f'{blockface_dir}volumes/raw/segmentation_volume.nrrd',
device=device)
# # Load the surface slice and get the difference
# blockface_surf = io.LoadITKFile(f'{blockface_dir}volumes/raw/surface/IMG_{bf_slice:03d}_surface.mhd',
# device=device)
#
# blockface_surf_p1 = io.LoadITKFile(f'{blockface_dir}volumes/raw/surface/IMG_{bf_slice + 1:03d}_surface.mhd',
# device=device)
#
# diff = (blockface_surf - blockface_surf_p1).data[2]
#
# diff = (diff - diff.min()) / (diff.max() - diff.min())
# Extract the slice
blockface_seg = blockface_seg.extract_slice(bf_slice - 1, dim=0)
aff_seg, affine = solve_affine(histology_seg,
blockface_seg,
img_num,
out_dir=out_dir,
device=device)
np.savetxt(f'{out_dir}/img_{img_num}_affine_to_blockface.txt',
affine.cpu().numpy())
#### Apply the affine to the image
mic_file = f'{histology_dir}hdf5/{block}_img{img_num}_image.hdf5'
meta_dict = {}
with h5py.File(mic_file, 'r') as f:
mic = f['RawImage/ImageData'][:, ::10, ::10]
for key in f['RawImage'].attrs:
meta_dict[key] = f['RawImage'].attrs[key]
mic = core.StructuredGrid(mic.shape[1:],
tensor=torch.tensor(mic,
dtype=torch.float32,
device=device),
spacing=torch.tensor([10.0, 10.0],
dtype=torch.float32,
device=device),
origin=histology_seg.origin,
device=device,
dtype=torch.float32,
channels=3)
mic = (mic - mic.min()) / (mic.max() - mic.min())
aff_mic = so.AffineTransform.Create(affine=affine)(mic, blockface_seg)
# plt.figure()
# plt.imshow(aff_mic.data.permute(1,2,0).cpu())
# plt.axis('off')
# plt.gca().invert_yaxis()
# plt.savefig(f'/home/sci/blakez/ucair/Animations/Scrolls/Mic/Images/{blockface_slice}_image.png', dpi=500, bbox_inches='tight', pad_inches=0)
def_histology, deformation = deformable_histology_to_blockface(
aff_seg,
blockface_seg,
steps=[0.01, 0.005],
scales=[4, 1],
gauss=True,
mic=aff_mic)
# Save out the deformation
io.SaveITKFile(deformation,
f'{out_dir}/img_{img_num}_deformation_to_blockface.mhd')
def solve_affines(block_list):
for block_path in block_list:
block = block_path.split('/')[-1]
if not os.path.exists(f'{block_path}/volumes/raw/'):
os.makedirs(f'{block_path}/volumes/raw/difference/')
os.makedirs(f'{block_path}/volumes/raw/surface/')
os.makedirs(f'{block_path}/volumes/raw/scatter/')
# elif sorted(glob.glob(f'{block_path}/volumes/raw/difference/*')):
# print('alread filtered')
# continue
image_list = sorted(glob.glob(f'{block_path}/images/filtered/*'))
if not image_list:
print(f'No filtered image files found for {block} ... skipping')
continue
print(f'Solving Affines for {block} ... ')
spacing = list(
map(float,
input(f'X,Y Spacing for {block}: ').strip().split(' ')))
# spacing = [0.0163, 0.0163]
surface_list = [x for x in image_list if 'surface' in x]
scatter_list = [x for x in image_list if 'scatter' in x]
ImScatter = io.LoadITKFile(scatter_list[0], device=device)
ImScatter.set_spacing_(spacing)
ImScatter.set_origin_(-1 * (ImScatter.size * ImScatter.spacing) / 2)
ImScatter /= ImScatter.max()
# Load the surface image
ImSurface = io.LoadITKFile(surface_list[0], device=device)
ImSurface.set_spacing_(spacing)
ImSurface.set_origin_(-1 * (ImSurface.size * ImSurface.spacing) / 2)
ImSurface /= ImSurface.max()
# Save out the first image
difference = ImScatter - ImSurface
ImDifference = core.StructuredGrid(
ImSurface.shape()[1:],
tensor=difference.data[2].unsqueeze(0),
spacing=ImSurface.spacing,
origin=ImSurface.origin,
device=device,
dtype=torch.float32,
channels=1)
io.SaveITKFile(
ImScatter, f'{block_path}/volumes/raw/scatter/IMG_001_scatter.mhd')
io.SaveITKFile(
ImSurface, f'{block_path}/volumes/raw/surface/IMG_001_surface.mhd')
io.SaveITKFile(
ImDifference,
f'{block_path}/volumes/raw/difference/IMG_001_difference.mhd')
ImPrev = ImScatter.copy()
Adict = {
'origin': ImPrev.origin.tolist(),
'spacing': ImPrev.spacing.tolist(),
scatter_list.index(scatter_list[0]): np.eye(3).tolist()
}
maxIter = 1000
for scat, surf in zip(scatter_list[1:], surface_list[1:]):
print(f'Registering {scat.split("/")[-1]} .... ')
sys.stdout.flush()
image_num = scat.split('/')[-1].split('_')[1]
# Get the number the file is from the start
dist = scatter_list.index(scat)
# Load the next image
ImScatter = io.LoadITKFile(scat, device=device)
ImScatter.set_spacing_(ImPrev.spacing)
ImScatter.set_origin_(ImPrev.origin)
ImScatter /= ImScatter.max()
ImSurface = io.LoadITKFile(surf, device=device)
ImSurface.set_spacing_(ImPrev.spacing)
ImSurface.set_origin_(ImPrev.origin)
ImSurface /= ImSurface.max()
difference = ImScatter - ImSurface
ImDifference = core.StructuredGrid(
ImSurface.shape()[1:],
tensor=difference.data[2].unsqueeze(0),
spacing=ImSurface.spacing,
origin=ImSurface.origin,
device=device,
dtype=torch.float32,
channels=1)
affine = affine_register(ImPrev.copy(),
ImSurface.copy(),
niter=maxIter,
device=device)
# Save out the images
aff_filter = so.AffineTransform.Create(affine=affine,
device=device)
aff_scatter = aff_filter(ImScatter)
aff_surface = aff_filter(ImSurface)
aff_difference = aff_filter(ImDifference)
# difference = (difference - difference.min()) / (difference.max() - difference.min())
io.SaveITKFile(
aff_scatter,
f'{block_path}/volumes/raw/scatter/IMG_{image_num}_scatter.mhd'
)
io.SaveITKFile(
aff_surface,
f'{block_path}/volumes/raw/surface/IMG_{image_num}_surface.mhd'
)
io.SaveITKFile(
aff_difference,
f'{block_path}/volumes/raw/difference/IMG_{image_num}_difference.mhd'
)
Adict[dist] = affine.detach().cpu().clone().tolist()
ImPrev = aff_scatter.copy()
with open(f'{block_path}/volumes/raw/{block}_affine_dictionary.yaml',
'w') as f:
yaml.dump(Adict, f)
_make_volumes(block_path)
def stitch_surfaces(rabbit):
rabbit_dir = f'/hdscratch/ucair/{rabbit}/blockface/'
raw_ext = '/surfaces/raw/'
vol_ext = '/volumes/raw/'
# Get a list of the blocks
block_list = sorted(glob.glob(f'{rabbit_dir}block*'))
# complete = ['block08']
complete = ['block06', 'block09']
for i, block_path in enumerate(block_list):
block = block_path.split('/')[-1]
stitching_dir = f'{rabbit_dir}{block}{raw_ext}/stitching/'
if block in complete:
continue
if not os.path.exists(stitching_dir):
print(f'No stitching surfaces found for {block}.')
continue
target_surface_path = f'{stitching_dir}/raw/{block}_target_faces.obj'
source_surface_path = f'{stitching_dir}/raw/{block}_source_faces.obj'
# Load the target surface
try:
verts, faces = io.ReadOBJ(target_surface_path)
tar_surface = core.TriangleMesh(verts, faces)
tar_surface.to_(device)
except IOError:
print(
f'The target stitching surface for {block} was not found ... skipping'
)
continue
try:
verts, faces = io.ReadOBJ(source_surface_path)
src_surface = core.TriangleMesh(verts, faces)
src_surface.to_(device)
src_surface.flip_normals_()
except IOError:
print(
f'The source stitching surface for {block} was not found ... skipping'
)
continue
# Need to load the exterior to drag along
try:
verts, faces = io.ReadOBJ(
f'{rabbit_dir}/{block}/{raw_ext}/{block}_decimate.obj')
surface_ext = core.TriangleMesh(verts, faces)
surface_ext.to_(device)
except IOError:
print(
f'The source stitching surface for {block} was not found ... skipping'
)
continue
# Determine the surface half way between the source and the target
try:
with open(
f'{stitching_dir}/raw/{block}_middle_surface_config.yaml',
'r') as f:
params = yaml.load(f, Loader=yaml.FullLoader)
except IOError:
params = {
'currents_sigma': [5.0, 0.25, 0.05],
'propagation_sigma': [1.0, 1.0, 1.0],
'deformable_lr': [0.0008, 0.01, 0.01],
'converge': 0.05,
'mid_offset': 0.5
}
# Do the deformable registration
def_src_surface, def_ext = tools.deformable_register(
tar_surface.copy(),
src_surface.copy(),
src_excess=None,
deformable_lr=params['deformable_lr'],
currents_sigma=params['currents_sigma'],
prop_sigma=params['propagation_sigma'],
grid_size=None,
converge=params['converge'],
accu_forward=False,
accu_inverse=False,
device=device,
)
new_verts = src_surface.vertices.clone() + (
(def_src_surface.vertices - src_surface.vertices) *
params['mid_offset'])
mid_surface = src_surface.copy()
mid_surface.vertices = new_verts.clone()
mid_surface.calc_normals()
mid_surface.calc_centers()
io.WriteOBJ(
def_src_surface.vertices, def_src_surface.indices,
f'{stitching_dir}/deformable_pieces/{block}_source_faces.obj')
io.WriteOBJ(
mid_surface.vertices, mid_surface.indices,
f'{stitching_dir}/deformable_pieces/{block}_source_middle.obj')
with open(f'{stitching_dir}/raw/{block}_middle_surface_config.yaml',
'w') as f:
yaml.dump(params, f)
# Load the binary volume for the block
mask = io.LoadITKFile(
f'{rabbit_dir}/{block}/{vol_ext}/segmentation_volume.mhd',
device='cuda:0')
# Load the other surfaces to drag along
extras_paths = [
f'{rabbit_dir}{block}{raw_ext}{block}_decimate.obj',
f'{rabbit_dir}{block}{raw_ext}{block}_ext.obj'
]
if os.path.exists(f'{rabbit_dir}{block}{raw_ext}{block}_head.obj'):
extras_paths += [f'{rabbit_dir}{block}{raw_ext}{block}_head.obj']
if os.path.exists(f'{rabbit_dir}{block}{raw_ext}{block}_foot.obj'):
extras_paths += [f'{rabbit_dir}{block}{raw_ext}{block}_foot.obj']
if os.path.exists(
f'{rabbit_dir}{block}{raw_ext}{block}_head_support.obj'):
extras_paths += [
f'{rabbit_dir}{block}{raw_ext}{block}_head_support.obj'
]
if os.path.exists(
f'{rabbit_dir}{block}{raw_ext}{block}_foot_support.obj'):
extras_paths += [
f'{rabbit_dir}{block}{raw_ext}{block}_foot_support.obj'
]
extra_surfaces = []
for path in extras_paths:
try:
verts, faces = io.ReadOBJ(path)
except IOError:
extra_name = path.split('/')[-1]
print(
f'{extra_name} not found as an extra ... removing from list'
)
_ = extras_paths.pop(extras_paths.index(path))
continue
extra_surfaces += [core.TriangleMesh(verts, faces)]
extra_surfaces[-1].to_(device)
# Define the diffusion parameters for the registration
try:
with open(f'{stitching_dir}/raw/{block}_diffusion_config.yaml',
'r') as f:
diff_params = yaml.load(f, Loader=yaml.FullLoader)
except IOError:
diff_params = {
'z_diff_c': 5.0,
'y_diff_c': 200.0,
'x_diff_c': 30.0,
'background_c': 0.5,
'niter': 10000,
'gamma': 0.0005,
'propegation_sigma': [0.2, 0.2, 0.2],
'grad_amp': 300.0
}
try:
with open(
f'{stitching_dir}/raw/{block}_stitch_surface_config.yaml',
'r') as f:
params = yaml.load(f, Loader=yaml.FullLoader)
except IOError:
params = {
'currents_sigma': [2.0, 1.0],
'smoothing_sigma': [2.0, 2.0, 2.0],
'deformable_lr': [0.0001, 0.001],
'grid_size': [40, 256, 256],
'converge': 0.05,
'grid_device': 'cpu',
'niters': 25
}
# Do the deformable registration with the source to the mid
def_src_surface, def_tar_surface, def_extra_surfaces, phi, phi_inv = deformable_register(
tar_surface.copy(),
src_surface.copy(),
mid_surface.copy(),
src_excess=extra_surfaces,
deformable_lr=params['deformable_lr'],
currents_sigma=params['currents_sigma'],
prop_sigma=params['smoothing_sigma'],
grid_size=params['grid_size'],
converge=params['converge'],
accu_forward=True,
accu_inverse=True,
grid_device=params['grid_device'],
device=device,
mask=mask,
diff_params=diff_params,
iters=params['niters'])
with open(f'{stitching_dir}/raw/{block}_stitch_surface_config.yaml',
'w') as f:
yaml.dump(params, f)
with open(f'{stitching_dir}/raw/{block}_diffusion_config.yaml',
'w') as f:
yaml.dump(diff_params, f)
io.WriteOBJ(
def_src_surface.vertices, def_src_surface.indices,
f'{stitching_dir}/deformable/{block}_whole_stitched_decimate.obj')
io.SaveITKFile(phi,
f'{stitching_dir}/deformable/{block}_stitch_phi.mhd')
io.SaveITKFile(
phi_inv, f'{stitching_dir}/deformable/{block}_stitch_phi_inv.mhd')
out_path = f'{stitching_dir}/deformable/'
for extra_path, extra_surface in zip(extras_paths, def_extra_surfaces):
name = extra_path.split('/')[-1]
io.WriteOBJ(extra_surface.vertices, extra_surface.indices,
f'{out_path}{name}')
vol = io.LoadITKFile(
'/hdscratch/ucair/18_047/blockface/block08/volumes/raw/difference_volume.mhd',
device='cuda:0')
# phi_inv.set_size((60, 1024, 1024))
# phi.set_size((60, 1024, 1024))
# resampled_stitched = so.ApplyGrid.Create(phi_inv, device='cuda:0')(vol, phi_inv)
# resampled_unstitched = so.ApplyGrid.Create(phi, device='cuda:0')(resampled_stitched, phi)
# io.SaveITKFile(resampled_stitched, '/home/sci/blakez/stitched_block08.mhd')
# io.SaveITKFile(resampled_unstitched, '/home/sci/blakez/unstitched_block08.mhd')
print(f'Done stitching {block} ... ')
def LoadDICOM(dicomDirectory, device):
'''Takes a directory that contains DICOM files and returns a PyCA Image3D
dicomDirectory = Full path to the folder with the dicoms
Returns an Image3D in the Reference Coordiante System (RCS)
'''
# Read the DICOM files in the directory
if type(dicomDirectory) == list:
dicoms = dicomDirectory
else:
dicoms = dc.read_dicom_directory(dicomDirectory)
# Sort the loaded dicoms
sort_dcms = dc.sort_dicoms(dicoms)
acqList = [sort_dcms[x].AcquisitionNumber for x in range(len(sort_dcms))]
# Need to account for if there are multiple acquisitions in a single folder
if np.min(acqList) != np.max(acqList):
volList = []
for acq in range(np.max(acqList)):
dcm_list = [x for x in sort_dcms if x.AcquisitionNumber == (acq + 1)]
volList += [LoadDICOM(dcm_list, device)]
return volList
# Extract the actual volume of pixels
pixel_vol = dc.get_volume_pixeldata(sort_dcms)
# Generate the affine from the dicom headers (THIS CODE WAS MODIFIED FROM dicom2nifti)
affine, spacing, pp = dc.create_affine(sort_dcms)
if type(affine) == str:
return 'Not A Volume'
# Convert the dicom volume to an Image3D - numpy is X, Y, Z and CAMP is Z, Y, X
rawDicom = core.StructuredGrid(
size=pixel_vol.shape,
origin=[0.0, 0.0, 0.0],
tensor=torch.tensor(pixel_vol.astype(np.float)).unsqueeze(0),
device=device,
channels=1
)
rcs_grid = SolveAffineGrid(rawDicom, affine)
wcsTrans = np.eye(4)
if pp == 'HFS':
wcsTrans[0, 0] *= -1
wcsTrans[1, 1] *= -1
if pp == 'FFS':
wcsTrans[1, 1] *= -1
wcsTrans[2, 2] *= -1
if pp == 'FFP':
wcsTrans[0, 0] *= -1
wcsTrans[2, 2] *= -1
world_grid = SolveAffineGrid(rcs_grid, wcsTrans)
rcs_grid = so.AffineTransform.Create(affine=torch.tensor(affine, dtype=torch.float, device=device))(
rawDicom, rcs_grid, xyz_affine=False)
world_grid = so.AffineTransform.Create(affine=torch.tensor(wcsTrans, dtype=torch.float, device=device))(
rcs_grid, world_grid, xyz_affine=False)
return world_grid
def process_mic(rabbit):
raw_mic_dir = f'/hdscratch/ucair/{rabbit}/microscopic/'
bf_dir = f'/hdscratch/ucair/{rabbit}/blockface/'
raw_bf_dir = f'/hdscratch/ucair/blockface/{rabbit}/'
from Histology.NNSeg.models import UNet
# from types import SimpleNamespace
import torch.nn as nn
from skimage import color
model = UNet.UNet(6, 3)
# saved_dict = SimpleNamespace(**torch.load(f'./Histology/NNSeg/model_weights/epoch_00230_model.pth'))
# if opt.cuda:
device = torch.device('cuda')
model = model.to(device=device)
model = nn.DataParallel(model)
params = torch.load('./NNSeg/model_weights/epoch_00230_model.pth')
model.load_state_dict(params['state_dict'])
# else:
# device = torch.device('cpu')
# params = torch.load(f'./Histology/NNSeg/model_weights/epoch_00230_model.pth', map_location='cpu')
# model.load_state_dict(params['state_dict'])
block_list = sorted(glob.glob(f'{raw_mic_dir}/block*'))
for block_path in block_list:
block = block_path.split('/')[-1]
# Make sure that the hdf5 files have been written
if not os.path.exists(f'{raw_mic_dir}{block}/hdf5/'):
os.makedirs(f'{raw_mic_dir}{block}/hdf5/')
convert_hdf5(block_path, raw_mic_dir)
mic_list = sorted(glob.glob(f'{block_path}/raw/*_image.tif'))
# if mic_list == []:
mic_list += sorted(glob.glob(f'{block_path}/raw/*_image.jpg'))
img_nums = [x.split('/')[-1].split('_')[1] for x in mic_list]
for img in img_nums:
if os.path.exists(
f'{raw_mic_dir}{block}/segmentations/IMG_{img}/img_{img}_gmm_segmentation.nii.gz'
):
continue
mic_file = f'{raw_mic_dir}{block}/hdf5/{block}_img{img}_image.hdf5'
meta_dict = {}
with h5py.File(mic_file, 'r') as f:
mic = f['RawImage/ImageData'][:, ::10, ::10]
for key in f['RawImage'].attrs:
meta_dict[key] = f['RawImage'].attrs[key]
mic = core.StructuredGrid(mic.shape[1:],
tensor=torch.tensor(mic,
dtype=torch.float32,
device=device),
spacing=torch.tensor([10.0, 10.0],
dtype=torch.float32,
device=device),
device='cpu',
dtype=torch.float32,
channels=3)
mic = (mic - mic.min()) / (mic.max() - mic.min())
plt.figure()
plt.imshow(mic.data.permute(1, 2, 0).cpu())
plt.title('Microscopic Image')
plt.show()
plt.pause(1.0)
satisfied = False
while not satisfied:
n_comps = int(
input("Enter the number of components for segmentation: "))
print('Clustering ... ', end='')
cluster = cluster_image(mic, n_comps)
print('done')
plt.figure()
plt.imshow(cluster)
plt.title('GMM Cluster')
plt.colorbar()
plt.show()
plt.pause(1.0)
redo_cluster = input(
"Are you satisfied with the clustering? [y/n]: ")
if redo_cluster == 'y':
satisfied = True
input_hsv = torch.from_numpy(
color.rgb2hsv(mic.data.squeeze().permute(1, 2,
0))).permute(2, 0, 1)
inputs = torch.cat([mic.data.squeeze(), input_hsv], dim=0)
pred = model(inputs.unsqueeze(0)).cpu().detach().squeeze()
seg = torch.zeros((pred.shape[1], pred.shape[2]))
for i in range(0, pred.shape[0]):
label = pred[i]
seg[label > 0] = i
# Save out the original image and the segmentation
out_path = f'{raw_mic_dir}{block}/segmentations/IMG_{img}/'
if not os.path.exists(out_path):
os.makedirs(out_path)
io.SaveITKFile(mic, f'{out_path}/img_{img}_color.nii.gz')
io.SaveITKFile(
core.StructuredGrid.FromGrid(mic, mic.data[0].unsqueeze(0)),
f'{out_path}/img_{img}_red.nii.gz')
io.SaveITKFile(
core.StructuredGrid.FromGrid(mic, mic.data[1].unsqueeze(0)),
f'{out_path}/img_{img}_green.nii.gz')
io.SaveITKFile(
core.StructuredGrid.FromGrid(mic, mic.data[2].unsqueeze(0)),
f'{out_path}/img_{img}_blue.nii.gz')
io.SaveITKFile(
core.StructuredGrid.FromGrid(
mic,
torch.tensor(cluster).unsqueeze(0)),
f'{out_path}/img_{img}_gmm_segmentation.nii.gz')
io.SaveITKFile(core.StructuredGrid.FromGrid(mic, seg.unsqueeze(0)),
f'{out_path}/img_nn_seg.nii.gz')
plt.close('all')
print(f'Done with {img}')
def register(rabbit, base_dir='/hdscratch/ucair/'):
source_path = f'{base_dir}{rabbit}/mri/exvivo/surfaces/raw/'
target_path = f'{base_dir}{rabbit}/mri/invivo/surfaces/raw/'
source_file = f'{source_path}exvivo_ablation_region_decimate.obj'
target_file = f'{target_path}invivo_ablation_region_decimate.obj'
verts, faces = io.ReadOBJ(target_file)
invivo_surface = core.TriangleMesh(verts, faces)
invivo_surface.to_(device=device)
verts, faces = io.ReadOBJ(source_file)
exvivo_surface = core.TriangleMesh(verts, faces)
exvivo_surface.to_(device=device)
print('Starting Affine ... ')
# Load or create the dictionary for registration
try:
with open(f'{source_path}affine_config.yaml', 'r') as f:
params = yaml.load(f, Loader=yaml.FullLoader)
except IOError:
params = {
'spatial_sigma': [15.0],
'affine_lr': 1.0e-07,
'translation_lr': 1.0e-06,
'converge': 4.0
}
try:
aff = np.loadtxt(f'{source_path}exvivo_to_invivo_affine.txt')
aff = torch.tensor(aff, device=device)
except IOError:
aff = affine(invivo_surface.copy(),
exvivo_surface.copy(),
affine_lr=params['affine_lr'],
translation_lr=params['translation_lr'],
converge=params['converge'],
spatial_sigma=params['spatial_sigma'],
device=device)
# Save out the parameters:
with open(f'{source_path}affine_config.yaml', 'w') as f:
yaml.dump(params, f)
np.savetxt(f'{source_path}exvivo_to_invivo_affine.txt',
aff.cpu().numpy())
aff_tfrom = uo.AffineTransformSurface.Create(aff, device=device)
aff_exvivo = aff_tfrom(exvivo_surface)
if not os.path.exists(f'{source_path}../affine/'):
os.makedirs(f'{source_path}../affine/')
io.WriteOBJ(
aff_exvivo.vertices, aff_exvivo.indices,
f'{source_path}../affine/exvivo_to_invivo_{rabbit}_affine.obj')
print('Starting Deformable ... ')
try:
with open(f'{source_path}/deformable_config.yaml', 'r') as f:
params = yaml.load(f, Loader=yaml.FullLoader)
except IOError:
params = {
'currents_sigma': [5.0, 1.0],
'propagation_sigma': [8.0, 8.0, 8.0],
'deformable_lr': [2.0e-04, 1.0e-04],
'converge': 4.0,
'phi_inv_size': [32, 32, 32],
'n_iters': 500,
}
def_surface, _, phi, phi_inv = tools.deformable_register(
invivo_surface.copy(),
aff_exvivo.copy(),
src_excess=None,
deformable_lr=params['deformable_lr'],
currents_sigma=params['currents_sigma'],
prop_sigma=params['propagation_sigma'],
converge=params['converge'],
grid_size=params['phi_inv_size'],
accu_forward=True,
accu_inverse=True,
device=device,
grid_device='cuda:0',
expansion_factor=1.5,
iters=params['n_iters'])
# Save out the parameters:
with open(f'{source_path}deformable_config.yaml', 'w') as f:
yaml.dump(params, f)
if not os.path.exists(f'{source_path}../../volumes/raw//'):
os.makedirs(f'{source_path}../../volumes/raw//')
if not os.path.exists(f'{source_path}../deformable/'):
os.makedirs(f'{source_path}../deformable/')
io.SaveITKFile(
phi_inv,
f'{source_path}../../volumes/raw/exvivo_to_invivo_phi_inv.mhd')
io.SaveITKFile(phi,
f'{source_path}../../volumes/raw/exvivo_to_invivo_phi.mhd')
io.WriteOBJ(def_surface.vertices, def_surface.indices,
f'{source_path}../deformable/exvivo_to_invivo_deformable.obj')