def preprocess_and_convert_to_numpy(nifti_scan: nib.Nifti1Image,
nifti_mask: nib.Nifti1Image) -> list:
""" Convert scan and label to numpy arrays and perform preprocessing
Return: Tuple(np.array, np.array) """
np_scan = nifti_scan.get_fdata()
np_label = nifti_mask.get_fdata()
nifti_mask.uncache()
nifti_scan.uncache()
np_scan = preprocess_scan(np_scan)
np_label = rotate_label(np_label)
assert np_scan.shape == np_label.shape
return np_scan, np_label
def analyze_kidney(seg:Nifti1Image,align=False)->list:
'''每个肾脏在数组中的体素点分布范围
args:
seg:Nifti1Image 体素点标注
align:bool 两肾选框是否对其
return：
list:list[tuple(slice,N)...] 两肾在各个维度上的范围切片对象，N为维度
centers:ndarray 两肾质心坐标
'''
data=seg.get_fdata()
indexs=argwhere(data>0)
kmeans=KMeans(n_clusters=2,precompute_distances=True).fit(indexs)
if align:
ranges=[]
for center in kmeans.cluster_centers_:
d,w,h=center
d_min,d_max=d-1,d+1
w_min,w_max=w-1,w+1
h_min,h_max=h-1,h+1
while d_min >0:
if KINDEY not in data[d_min,:,:]:
break
d_min-=1
while w_min >0:
if KINDEY not in data[:,w_min,:]:
break
w_min-=1
while h_min >0:
if KINDEY not in data[:,:,h_min]:
break
h_min-=1
while d_max < data.shape[0]:
if KINDEY not in data[d_max,:,:]:
break
d_max+=1
while w_max < data.shape[1]:
if KINDEY not in data[:,w_max,:]:
break
w_max+=1
while h_max <data.shape[2]:
if KINDEY not in data[:,:,h_max]:
break
h_max+=1
ranges.append((slice(d_min,d_max),slice(w_min,w_max),slice(h_min,h_max)))
return ranges,array(kmeans.cluster_centers_,dtype=int)
else:
labels=kmeans.predict(indexs)
kindey1=indexs[labels==0]
kindey2=indexs[labels==1]
d_min_1,w_min_1,h_min_1=kindey1.min(axis=0)
d_max_1,w_max_1,h_max_1=kindey1.max(axis=0)
d_min_2,w_min_2,h_min_2=kindey2.min(axis=0)
d_max_2,w_max_2,h_max_2=kindey2.max(axis=0)
return [
(slice(d_min_1,d_max_1),slice(w_min_1,w_max_1),slice(h_min_1,h_max_1)),
(slice(d_min_2,d_max_2),slice(w_min_2,w_max_2),slice(h_min_2,h_max_2))
],array(kmeans.cluster_centers_,dtype=int)
def decompose(img: nib.Nifti1Image, name: str, orientation: str, scale: int,
pad_crop: list):
"""
Splits a nibabel Nifti1Image into separate slices given orientation to make slices on
"""
# Force RAS+ orientation
#img = nib.as_closest_canonical(img)
# get orient code
orient_code = {'S': 0, 'C': 1, 'T': 2}[orientation]
# get volume data
data = img.get_fdata()
dims = data.shape
# scale data
data = np.round((data / scale) * 65535)
# write images to disk
for n in range(dims[orient_code]):
output_filename = name + '_slice-{:0>4d}.png'.format(n)
data_slice = np.flip(simple_slice(data, n,
orient_code).T.astype('uint16'),
axis=0)
if pad_crop: # do padding/crop if enabled
data_slice = resize_slice(data_slice, pad_crop)
# write slice to file
imwrite(output_filename, data_slice)
def _clear_steady_state(img: Nifti1Image,
confounds: pd.DataFrame,
drop_trs: Optional[int] = None):
"""
Remove steady state volumes
"""
if not drop_trs:
steady_cols = [c for c in confounds.columns if "steady" in c]
if steady_cols:
steady_df = confounds[steady_cols].sum(axis=1).diff()
steady_ind = np.where(steady_df < 0)[0]
drop_trs = int(steady_ind[0])
else:
raise ValueError(
"drop_trs not supplied and steady_state volumes not"
" found in confounds dataframe!")
# Construct new image object
new_conf = confounds.loc[drop_trs:, :]
new_img = nimg.new_img_like(img,
img.get_fdata(caching="unchanged")[:, :, :,
drop_trs:],
copy_header=True)
return (new_img, new_conf)
def _get_vol_index(img: Nifti1Image, inds: npt.ArrayLike) -> Nifti1Image:
return nimg.new_img_like(
img,
img.get_fdata(caching="unchanged")[:, :, :, inds],
img.affine,
copy_header=True,
)
def _image_to_signals(img: Nifti1Image) -> npt.ArrayLike:
"""
Transform a Nibabel image into an [NVOX x T] array
"""
nvox = np.prod(img.shape[:-1])
return img.get_fdata(caching="unchanged").reshape((nvox, -1))
def fix_mosaic(mosaic_nifti: nib.Nifti1Image, acq_dims: tuple):
"""
Fixes incorrectly-processed NIFTIs by dcm2niix where they still remain mosaics due to a lack of
NumberOfImagesInMosaic header. This function implements a hack to
:param mosaic_nifti: the nifti image object that needs to be fixed. Should be of shape m x n x 1
the sliding window algorithm
:param acq_dims: the (row, col) acquisition dimensions for rows and columns from the AcquisitionMatrix DICOM
field.
Used to determine the appropriate kernel size to use for the sliding window algorithm
:return: new_nifti; a 3D NIFTI that is no longer mosaic
"""
acq_rows, acq_cols = acq_dims
# Get the shape and array values of the mosaic (flatten the latter into a 2D array)
img_shape = mosaic_nifti.shape
# noinspection PyTypeChecker
img_data = np.rot90(np.squeeze(mosaic_nifti.get_fdata()))
# If this is a square, and the rows perfectly divides the mosaic
if img_shape[0] == img_shape[1] and img_shape[0] % acq_rows == 0:
nsplits_w, nsplits_h = img_shape[0] / acq_rows, img_shape[0] / acq_rows
kernel_w, kernel_h = acq_rows, acq_rows
# If this is a square, and the cols perfectly divides the mosaic
elif img_shape[0] == img_shape[1] and img_shape[0] % acq_cols == 0:
nsplits_w, nsplits_h = img_shape[0] / acq_cols, img_shape[0] / acq_cols
kernel_w, kernel_h = acq_cols, acq_cols
# If this is a rectangle
elif all([img_shape[0] != img_shape[1],
img_shape[0] % acq_rows == 0,
img_shape[1] % acq_cols == 0
]):
nsplits_w, nsplits_h = img_shape[0] / acq_rows, img_shape[1] / acq_cols
kernel_w, kernel_h = acq_rows, acq_cols
else:
return
# Initialize the data that will house the split mosaic into slices
new_img_data = np.zeros(shape=(kernel_w, kernel_h, int(nsplits_w * nsplits_h)))
slice_num = 0
# Sliding Window algorithm
for ii in range(int(nsplits_w)):
for jj in range(int(nsplits_h)):
x_start, x_end = ii * kernel_w, (ii + 1) * kernel_w
y_start, y_end = jj * kernel_h, (jj + 1) * kernel_h
img_slice = img_data[x_start:x_end, y_start:y_end]
# Disregard slices that are only zeros
if np.nanmax(img_slice) == 0:
continue
# Otherwise update the zeros array at the appropriate slice with the new values
else:
new_img_data[:, :, slice_num] = img_slice
slice_num += 1
# Filter off slices that had only zeros
new_img_data = np.rot90(new_img_data[:, :, 0:slice_num], 3)
new_nifti = image.new_img_like(mosaic_nifti, new_img_data, affine=mosaic_nifti.affine)
return new_nifti
def get_data(img:Nifti1Image)->ndarray:
'''获取影像数据
args：
img:Nifti1Image
return:
data:ndarray size[d,w,h]
'''
return img.get_fdata()
def analyze_cubes(seg:Nifti1Image)->list:
'''细分割（每种前景）体素点在数组中的分布范围
args:
seg:Nifti1Image 体素点标注
return：
list:list[tuple(slice,N)...] 每种前景对象在各个维度上的范围切片对象，N为维度
'''
return ndimage.find_objects(seg.get_fdata().astype(int))
def robust_set_limits_in_mask(data_img: nib.Nifti1Image,
mask_img: nib.Nifti1Image) -> dict[str, float]:
plot_params: dict[str, float] = dict()
mask = np.asanyarray(mask_img.dataobj).astype(bool)
data = data_img.get_fdata()[mask]
plot_params = robust_set_limits(data.reshape(-1), plot_params)
return plot_params
def analyze_cube(seg:Nifti1Image)->tuple:
'''粗分割（前景）体素点在数组中的分布范围
args:
seg:Nifti1Image 体素点标注
return：
slice_tuple:tuple(slice,N) 各个维度上的范围切片对象，N为维度
'''
indexs=argwhere(seg.get_fdata()!=0)
return tuple(slice(start,end+1) for start,end in zip(indexs.min(axis=0),indexs.max(axis=0)))
def analyze_kidney_center(seg:Nifti1Image)->list:
'''每个肾脏在3D图像中坐标的质心
args:
seg:Nifti1Image 体素点标注
return:
list:[[d,w,h],[d,w,h]]
'''
indexs=argwhere(seg.get_fdata()>0)
kmeans=KMeans(n_clusters=2,precompute_distances=True).fit(indexs)
return array(kmeans.cluster_centers_,dtype=int)
def from_image(cls, image: Nifti1Image, space: Space, ignore_affine=False):
"""Construct a bounding box from a nifti image"""
bounds = cls._determine_bounds(image.get_fdata())
if bounds is None:
return None
if ignore_affine:
target_space = None
else:
bounds = np.dot(image.affine, bounds)
target_space = space
return cls(point1=bounds[:3, 0], point2=bounds[:3, 1], space=target_space)
def analyze_mean_std(img:Nifti1Image):
'''分析影像体素点的均值和方差
args:
img:Nifti1Image 待分析影像
return：
mean:float 均值
std:float 方差
'''
data=img.get_fdata()
mean=data.mean()
std=data.std()
return mean,std
def resample_image(img:Nifti1Image,space_target:ndarray=array([3.22,1.62,1.62]))->Nifti1Image:
'''对影像数据进行重采样（类型为Nifti1Image）
args:
img:Nifti1Image 原始数组
space_target 目标间距
return：
data:Nifti1Image 重采样后的影像数据
'''
aff=-diag([*space_target,-1])
aff=matmul(array([
[0,0,1,0],
[0,1,0,0],
[1,0,0,0],
[0,0,0,1],
]),aff)
scales=get_spacing(img)/space_target
data=img.get_fdata()
resample=ndimage.zoom(data,scales,mode="reflect")
return Nifti1Image(resample,aff)
def resample_segmentation(seg:Nifti1Image,space_target:ndarray=array([3.22,1.62,1.62]))->Nifti1Image:
'''对分割标注数据进行重采样（类型为Nifti1Image）
args:
img:Nifti1Image 原始影像
space_target 目标间距
return：
data:Nifti1Image 重采样后的分割标注数据
'''
aff=-diag([*space_target,-1])
aff=matmul(array([
[0,0,1,0],
[0,1,0,0],
[1,0,0,0],
[0,0,0,1],
]),aff)
scales=get_spacing(seg)/space_target
data=seg.get_fdata()
resample=ndimage.interpolation.zoom(data,scales,order=1,mode="reflect")
resample=around(resample,0).astype(int)
return Nifti1Image(resample,aff)
def weight(echo: nib.Nifti1Image, TE: float):
data = echo.get_fdata()
mean = data[..., -n_vols:].mean(axis=-1)
std = data[..., -n_vols:].std(axis=-1)
return TE * mean / std
