AdritRao's picture
Upload 62 files
a3290d1
raw
history blame
No virus
25.6 kB
"""
@author: louisblankemeier
"""
import logging
import math
import os
from typing import Dict, List
import cv2
import nibabel as nib
import numpy as np
from scipy.ndimage import zoom
from comp2comp.spine import spine_visualization
def find_spine_dicoms(centroids: Dict): # , path: str, levels):
"""Find the dicom files corresponding to the spine T12 - L5 levels."""
vertical_positions = []
for level in centroids:
centroid = centroids[level]
vertical_positions.append(round(centroid[2]))
# dicom_files = []
# ipps = []
# for dicom_path in glob(path + "/*.dcm"):
# ipp = dcmread(dicom_path).ImagePositionPatient
# ipps.append(ipp[2])
# dicom_files.append(dicom_path)
# dicom_files = [x for _, x in sorted(zip(ipps, dicom_files))]
# dicom_files = list(np.array(dicom_files)[vertical_positions])
# return (dicom_files, levels, vertical_positions)
return vertical_positions
def save_nifti_select_slices(output_dir: str, vertical_positions):
nifti_path = os.path.join(output_dir, "segmentations", "converted_dcm.nii.gz")
nifti_in = nib.load(nifti_path)
nifti_np = nifti_in.get_fdata()
nifti_np = nifti_np[:, :, vertical_positions]
nifti_out = nib.Nifti1Image(nifti_np, nifti_in.affine, nifti_in.header)
# save the nifti
nifti_output_path = os.path.join(
output_dir, "segmentations", "converted_dcm.nii.gz"
)
nib.save(nifti_out, nifti_output_path)
# Function that takes a numpy array as input, computes the
# sagittal centroid of each label and returns a list of the
# centroids
def compute_centroids(seg: np.ndarray, spine_model_type):
"""Compute the centroids of the labels.
Args:
seg (np.ndarray): Segmentation volume.
spine_model_type (str): Model type.
Returns:
List[int]: List of centroids.
"""
# take values of spine_model_type.categories dictionary
# and convert to list
centroids = {}
for level in spine_model_type.categories:
label_idx = spine_model_type.categories[level]
try:
pos = compute_centroid(seg, "sagittal", label_idx)
centroids[level] = pos
except Exception:
logging.warning(f"Label {level} not found in segmentation volume.")
return centroids
# Function that takes a numpy array as input, as well as a list of centroids,
# takes a slice through the centroid on axis = 1 for each centroid
# and returns a list of the slices
def get_slices(seg: np.ndarray, centroids: Dict, spine_model_type):
"""Get the slices corresponding to the centroids.
Args:
seg (np.ndarray): Segmentation volume.
centroids (List[int]): List of centroids.
spine_model_type (str): Model type.
Returns:
List[np.ndarray]: List of slices.
"""
seg = seg.astype(np.uint8)
slices = {}
for level in centroids:
label_idx = spine_model_type.categories[level]
binary_seg = (seg[centroids[level], :, :] == label_idx).astype(int)
if (
np.sum(binary_seg) > 200
): # heuristic to make sure enough of the body is showing
slices[level] = binary_seg
return slices
# Function that takes a mask and for each deletes the right most
# connected component. Returns the mask with the right most
# connected component deleted
def delete_right_most_connected_component(mask: np.ndarray):
"""Delete the right most connected component corresponding to spinous processes.
Args:
mask (np.ndarray): Mask volume.
Returns:
np.ndarray: Mask volume.
"""
mask = mask.astype(np.uint8)
_, labels, _, centroids = cv2.connectedComponentsWithStats(mask, connectivity=8)
right_most_connected_component = np.argmin(centroids[1:, 1]) + 1
mask[labels == right_most_connected_component] = 0
return mask
# compute center of mass of 2d mask
def compute_center_of_mass(mask: np.ndarray):
"""Compute the center of mass of a 2D mask.
Args:
mask (np.ndarray): Mask volume.
Returns:
np.ndarray: Center of mass.
"""
mask = mask.astype(np.uint8)
_, _, _, centroids = cv2.connectedComponentsWithStats(mask, connectivity=8)
center_of_mass = np.mean(centroids[1:, :], axis=0)
return center_of_mass
# Function that takes a 3d centroid and retruns a binary mask with a 3d
# roi around the centroid
def roi_from_mask(img, centroid: np.ndarray):
"""Compute a 3D ROI from a 3D mask.
Args:
img (np.ndarray): Image volume.
centroid (np.ndarray): Centroid.
Returns:
np.ndarray: ROI volume.
"""
roi = np.zeros(img.shape)
img_np = img.get_fdata()
pixel_spacing = img.header.get_zooms()
length_i = 5.0 / pixel_spacing[0]
length_j = 5.0 / pixel_spacing[1]
length_k = 5.0 / pixel_spacing[2]
print(
f"Computing ROI with centroid {centroid[0]:.3f}, {centroid[1]:.3f}, {centroid[2]:.3f} "
f"and pixel spacing "
f"{pixel_spacing[0]:.3f}mm, {pixel_spacing[1]:.3f}mm, {pixel_spacing[2]:.3f}mm..."
)
# cubic ROI around centroid
"""
roi[
int(centroid[0] - length) : int(centroid[0] + length),
int(centroid[1] - length) : int(centroid[1] + length),
int(centroid[2] - length) : int(centroid[2] + length),
] = 1
"""
# spherical ROI around centroid
roi = np.zeros(img_np.shape)
i_lower = math.floor(centroid[0] - length_i)
j_lower = math.floor(centroid[1] - length_j)
k_lower = math.floor(centroid[2] - length_k)
i_lower_idx = 1000
j_lower_idx = 1000
k_lower_idx = 1000
i_upper_idx = 0
j_upper_idx = 0
k_upper_idx = 0
found_pixels = False
for i in range(i_lower, i_lower + 2 * math.ceil(length_i) + 1):
for j in range(j_lower, j_lower + 2 * math.ceil(length_j) + 1):
for k in range(k_lower, k_lower + 2 * math.ceil(length_k) + 1):
if (i - centroid[0]) ** 2 / length_i**2 + (
j - centroid[1]
) ** 2 / length_j**2 + (k - centroid[2]) ** 2 / length_k**2 <= 1:
roi[i, j, k] = 1
if i < i_lower_idx:
i_lower_idx = i
if j < j_lower_idx:
j_lower_idx = j
if k < k_lower_idx:
k_lower_idx = k
if i > i_upper_idx:
i_upper_idx = i
if j > j_upper_idx:
j_upper_idx = j
if k > k_upper_idx:
k_upper_idx = k
found_pixels = True
if not found_pixels:
print("No pixels in ROI!")
raise ValueError
print(
f"Number of pixels included in i, j, and k directions: {i_upper_idx - i_lower_idx + 1}, "
f"{j_upper_idx - j_lower_idx + 1}, {k_upper_idx - k_lower_idx + 1}"
)
return roi
# Function that takes a 3d image and a 3d binary mask and returns that average
# value of the image inside the mask
def mean_img_mask(img: np.ndarray, mask: np.ndarray, index: int):
"""Compute the mean of an image inside a mask.
Args:
img (np.ndarray): Image volume.
mask (np.ndarray): Mask volume.
rescale_slope (float): Rescale slope.
rescale_intercept (float): Rescale intercept.
Returns:
float: Mean value.
"""
img = img.astype(np.float32)
mask = mask.astype(np.float32)
img_masked = (img * mask)[mask > 0]
# mean = (rescale_slope * np.mean(img_masked)) + rescale_intercept
# median = (rescale_slope * np.median(img_masked)) + rescale_intercept
mean = np.mean(img_masked)
return mean
def compute_rois(seg, img, spine_model_type):
"""Compute the ROIs for the spine.
Args:
seg (np.ndarray): Segmentation volume.
img (np.ndarray): Image volume.
rescale_slope (float): Rescale slope.
rescale_intercept (float): Rescale intercept.
spine_model_type (Models): Model type.
Returns:
spine_hus (List[float]): List of HU values.
rois (List[np.ndarray]): List of ROIs.
centroids_3d (List[np.ndarray]): List of centroids.
"""
seg_np = seg.get_fdata()
centroids = compute_centroids(seg_np, spine_model_type)
slices = get_slices(seg_np, centroids, spine_model_type)
for level in slices:
slice = slices[level]
# keep only the two largest connected components
two_largest, two = keep_two_largest_connected_components(slice)
if two:
slices[level] = delete_right_most_connected_component(two_largest)
# Compute ROIs
rois = {}
spine_hus = {}
centroids_3d = {}
segmentation_hus = {}
for i, level in enumerate(slices):
slice = slices[level]
center_of_mass = compute_center_of_mass(slice)
centroid = np.array([centroids[level], center_of_mass[1], center_of_mass[0]])
roi = roi_from_mask(img, centroid)
image_numpy = img.get_fdata()
spine_hus[level] = mean_img_mask(image_numpy, roi, i)
rois[level] = roi
mask = (seg_np == spine_model_type.categories[level]).astype(int)
segmentation_hus[level] = mean_img_mask(image_numpy, mask, i)
centroids_3d[level] = centroid
return (spine_hus, rois, segmentation_hus, centroids_3d)
def keep_two_largest_connected_components(mask: Dict):
"""Keep the two largest connected components.
Args:
mask (np.ndarray): Mask volume.
Returns:
np.ndarray: Mask volume.
"""
mask = mask.astype(np.uint8)
# sort connected components by size
num_labels, labels, stats, centroids = cv2.connectedComponentsWithStats(
mask, connectivity=8
)
stats = stats[1:, 4]
sorted_indices = np.argsort(stats)[::-1]
# keep only the two largest connected components
mask = np.zeros(mask.shape)
mask[labels == sorted_indices[0] + 1] = 1
two = True
try:
mask[labels == sorted_indices[1] + 1] = 1
except Exception:
two = False
return (mask, two)
def compute_centroid(seg: np.ndarray, plane: str, label: int):
"""Compute the centroid of a label in a given plane.
Args:
seg (np.ndarray): Segmentation volume.
plane (str): Plane.
label (int): Label.
Returns:
int: Centroid.
"""
if plane == "axial":
sum_out_axes = (0, 1)
sum_axis = 2
elif plane == "sagittal":
sum_out_axes = (1, 2)
sum_axis = 0
elif plane == "coronal":
sum_out_axes = (0, 2)
sum_axis = 1
sums = np.sum(seg == label, axis=sum_out_axes)
normalized_sums = sums / np.sum(sums)
pos = int(np.sum(np.arange(0, seg.shape[sum_axis]) * normalized_sums))
return pos
def to_one_hot(label: np.ndarray, model_type, spine_hus):
"""Convert a label to one-hot encoding.
Args:
label (np.ndarray): Label volume.
model_type (Models): Model type.
Returns:
np.ndarray: One-hot encoding volume.
"""
levels = list(spine_hus.keys())
levels.reverse()
one_hot_label = np.zeros((label.shape[0], label.shape[1], len(levels)))
for i, level in enumerate(levels):
label_idx = model_type.categories[level]
one_hot_label[:, :, i] = (label == label_idx).astype(int)
return one_hot_label
def visualize_coronal_sagittal_spine(
seg: np.ndarray,
rois: List[np.ndarray],
mvs: np.ndarray,
centroids_3d: np.ndarray,
output_dir: str,
spine_hus=None,
seg_hus=None,
model_type=None,
pixel_spacing=None,
format="png",
):
"""Visualize the coronal and sagittal planes of the spine.
Args:
seg (np.ndarray): Segmentation volume.
rois (List[np.ndarray]): List of ROIs.
mvs (dm.MedicalVolume): Medical volume.
centroids (List[int]): List of centroids.
label_text (List[str]): List of labels.
output_dir (str): Output directory.
spine_hus (List[float], optional): List of HU values. Defaults to None.
model_type (Models, optional): Model type. Defaults to None.
"""
sagittal_vals, coronal_vals = curved_planar_reformation(mvs, centroids_3d)
zoom_factor = pixel_spacing[2] / pixel_spacing[1]
sagittal_image = mvs[sagittal_vals, :, range(len(sagittal_vals))]
sagittal_label = seg[sagittal_vals, :, range(len(sagittal_vals))]
sagittal_image = zoom(sagittal_image, (zoom_factor, 1), order=3)
sagittal_label = zoom(sagittal_label, (zoom_factor, 1), order=1).round()
one_hot_sag_label = to_one_hot(sagittal_label, model_type, spine_hus)
for roi in rois:
one_hot_roi_label = roi[sagittal_vals, :, range(len(sagittal_vals))]
one_hot_roi_label = zoom(one_hot_roi_label, (zoom_factor, 1), order=1).round()
one_hot_sag_label = np.concatenate(
(
one_hot_sag_label,
one_hot_roi_label.reshape(
(one_hot_roi_label.shape[0], one_hot_roi_label.shape[1], 1)
),
),
axis=2,
)
coronal_image = mvs[:, coronal_vals, range(len(coronal_vals))]
coronal_label = seg[:, coronal_vals, range(len(coronal_vals))]
coronal_image = zoom(coronal_image, (1, zoom_factor), order=3)
coronal_label = zoom(coronal_label, (1, zoom_factor), order=1).round()
# coronal_image = zoom(coronal_image, (zoom_factor, 1), order=3)
# coronal_label = zoom(coronal_label, (zoom_factor, 1), order=0).astype(int)
one_hot_cor_label = to_one_hot(coronal_label, model_type, spine_hus)
for roi in rois:
one_hot_roi_label = roi[:, coronal_vals, range(len(coronal_vals))]
one_hot_roi_label = zoom(one_hot_roi_label, (1, zoom_factor), order=1).round()
one_hot_cor_label = np.concatenate(
(
one_hot_cor_label,
one_hot_roi_label.reshape(
(one_hot_roi_label.shape[0], one_hot_roi_label.shape[1], 1)
),
),
axis=2,
)
# flip both axes of coronal image
sagittal_image = np.flip(sagittal_image, axis=0)
sagittal_image = np.flip(sagittal_image, axis=1)
# flip both axes of coronal label
one_hot_sag_label = np.flip(one_hot_sag_label, axis=0)
one_hot_sag_label = np.flip(one_hot_sag_label, axis=1)
coronal_image = np.transpose(coronal_image)
one_hot_cor_label = np.transpose(one_hot_cor_label, (1, 0, 2))
# flip both axes of sagittal image
coronal_image = np.flip(coronal_image, axis=0)
coronal_image = np.flip(coronal_image, axis=1)
# flip both axes of sagittal label
one_hot_cor_label = np.flip(one_hot_cor_label, axis=0)
one_hot_cor_label = np.flip(one_hot_cor_label, axis=1)
if format == "png":
sagittal_name = "spine_sagittal.png"
coronal_name = "spine_coronal.png"
elif format == "dcm":
sagittal_name = "spine_sagittal.dcm"
coronal_name = "spine_coronal.dcm"
else:
raise ValueError("Format must be either png or dcm")
img_sagittal = spine_visualization.spine_binary_segmentation_overlay(
sagittal_image,
one_hot_sag_label,
output_dir,
sagittal_name,
spine_hus=spine_hus,
seg_hus=seg_hus,
model_type=model_type,
pixel_spacing=pixel_spacing,
)
img_coronal = spine_visualization.spine_binary_segmentation_overlay(
coronal_image,
one_hot_cor_label,
output_dir,
coronal_name,
spine_hus=spine_hus,
seg_hus=seg_hus,
model_type=model_type,
pixel_spacing=pixel_spacing,
)
return img_sagittal, img_coronal
def curved_planar_reformation(mvs, centroids):
centroids = sorted(centroids, key=lambda x: x[2])
centroids = [(int(x[0]), int(x[1]), int(x[2])) for x in centroids]
sagittal_centroids = [centroids[i][0] for i in range(0, len(centroids))]
coronal_centroids = [centroids[i][1] for i in range(0, len(centroids))]
axial_centroids = [centroids[i][2] for i in range(0, len(centroids))]
sagittal_vals = [sagittal_centroids[0]] * axial_centroids[0]
coronal_vals = [coronal_centroids[0]] * axial_centroids[0]
for i in range(1, len(axial_centroids)):
num = axial_centroids[i] - axial_centroids[i - 1]
interp = list(
np.linspace(sagittal_centroids[i - 1], sagittal_centroids[i], num=num)
)
sagittal_vals.extend(interp)
interp = list(
np.linspace(coronal_centroids[i - 1], coronal_centroids[i], num=num)
)
coronal_vals.extend(interp)
sagittal_vals.extend([sagittal_centroids[-1]] * (mvs.shape[2] - len(sagittal_vals)))
coronal_vals.extend([coronal_centroids[-1]] * (mvs.shape[2] - len(coronal_vals)))
sagittal_vals = np.array(sagittal_vals)
coronal_vals = np.array(coronal_vals)
sagittal_vals = sagittal_vals.astype(int)
coronal_vals = coronal_vals.astype(int)
return (sagittal_vals, coronal_vals)
'''
def compare_ts_stanford_centroids(labels_path, pred_centroids):
"""Compare the centroids of the Stanford dataset with the centroids of the TS dataset.
Args:
labels_path (str): Path to the Stanford dataset labels.
"""
t12_diff = []
l1_diff = []
l2_diff = []
l3_diff = []
l4_diff = []
l5_diff = []
num_skipped = 0
labels = glob(labels_path + "/*")
for label_path in labels:
# modify label_path to give pred_path
pred_path = label_path.replace("labelsTs", "predTs_TS")
print(label_path.split("/")[-1])
label_nib = nib.load(label_path)
label = label_nib.get_fdata()
spacing = label_nib.header.get_zooms()[2]
pred_nib = nib.load(pred_path)
pred = pred_nib.get_fdata()
if True:
pred[pred == 18] = 6
pred[pred == 19] = 5
pred[pred == 20] = 4
pred[pred == 21] = 3
pred[pred == 22] = 2
pred[pred == 23] = 1
for label_idx in range(1, 7):
label_level = label == label_idx
indexes = np.array(range(label.shape[2]))
sums = np.sum(label_level, axis=(0, 1))
normalized_sums = sums / np.sum(sums)
label_centroid = np.sum(indexes * normalized_sums)
print(f"Centroid for label {label_idx}: {label_centroid}")
if False:
try:
pred_centroid = pred_centroids[6 - label_idx]
except Exception:
# Change this part
print("Something wrong with pred_centroids, skipping!")
num_skipped += 1
break
# if revert_to_original:
if True:
pred_level = pred == label_idx
sums = np.sum(pred_level, axis=(0, 1))
indices = list(range(sums.shape[0]))
groupby_input = zip(indices, list(sums))
g = groupby(groupby_input, key=lambda x: x[1] > 0.0)
m = max([list(s) for v, s in g if v > 0], key=lambda x: np.sum(list(zip(*x))[1]))
res = list(zip(*m))
indexes = list(res[0])
sums = list(res[1])
normalized_sums = sums / np.sum(sums)
pred_centroid = np.sum(indexes * normalized_sums)
print(f"Centroid for prediction {label_idx}: {pred_centroid}")
diff = np.absolute(pred_centroid - label_centroid) * spacing
if label_idx == 1:
t12_diff.append(diff)
elif label_idx == 2:
l1_diff.append(diff)
elif label_idx == 3:
l2_diff.append(diff)
elif label_idx == 4:
l3_diff.append(diff)
elif label_idx == 5:
l4_diff.append(diff)
elif label_idx == 6:
l5_diff.append(diff)
print(f"Skipped {num_skipped}")
print("The final mean differences in mm:")
print(
np.mean(t12_diff),
np.mean(l1_diff),
np.mean(l2_diff),
np.mean(l3_diff),
np.mean(l4_diff),
np.mean(l5_diff),
)
print("The final median differences in mm:")
print(
np.median(t12_diff),
np.median(l1_diff),
np.median(l2_diff),
np.median(l3_diff),
np.median(l4_diff),
np.median(l5_diff),
)
def compare_ts_stanford_roi_hus(image_path):
"""Compare the HU values of the Stanford dataset with the HU values of the TS dataset.
image_path (str): Path to the Stanford dataset images.
"""
img_paths = glob(image_path + "/*")
differences = np.zeros((40, 6))
ground_truth = np.zeros((40, 6))
for i, img_path in enumerate(img_paths):
print(f"Image number {i + 1}")
image_path_no_0000 = re.sub(r"_0000", "", img_path)
ts_seg_path = image_path_no_0000.replace("imagesTs", "predTs_TS")
stanford_seg_path = image_path_no_0000.replace("imagesTs", "labelsTs")
img = nib.load(img_path).get_fdata()
img = np.swapaxes(img, 0, 1)
ts_seg = nib.load(ts_seg_path).get_fdata()
ts_seg = np.swapaxes(ts_seg, 0, 1)
stanford_seg = nib.load(stanford_seg_path).get_fdata()
stanford_seg = np.swapaxes(stanford_seg, 0, 1)
ts_model_type = Models.model_from_name("ts_spine")
(spine_hus_ts, rois, centroids_3d) = compute_rois(ts_seg, img, 1, 0, ts_model_type)
stanford_model_type = Models.model_from_name("stanford_spine_v0.0.1")
(spine_hus_stanford, rois, centroids_3d) = compute_rois(
stanford_seg, img, 1, 0, stanford_model_type
)
difference_vals = np.abs(np.array(spine_hus_ts) - np.array(spine_hus_stanford))
print(f"Differences {difference_vals}\n")
differences[i, :] = difference_vals
ground_truth[i, :] = spine_hus_stanford
print("\n")
# compute average percent change from ground truth
percent_change = np.divide(differences, ground_truth) * 100
average_percent_change = np.mean(percent_change, axis=0)
median_percent_change = np.median(percent_change, axis=0)
# print average percent change
print("Average percent change from ground truth:")
print(average_percent_change)
print("Median percent change from ground truth:")
print(median_percent_change)
# print average difference
average_difference = np.mean(differences, axis=0)
median_difference = np.median(differences, axis=0)
print("Average difference from ground truth:")
print(average_difference)
print("Median difference from ground truth:")
print(median_difference)
def process_post_hoc(pred_path):
"""Apply post-hoc heuristics for improving Stanford spine model vertical centroid predictions.
Args:
pred_path (str): Path to the prediction.
"""
pred_nib = nib.load(pred_path)
pred = pred_nib.get_fdata()
pred_bodies = np.logical_and(pred >= 1, pred <= 6)
pred_bodies = pred_bodies.astype(np.int64)
labels_out, N = cc3d.connected_components(pred_bodies, return_N=True, connectivity=6)
stats = cc3d.statistics(labels_out)
print(stats)
labels_out_list = []
voxel_counts_list = list(stats["voxel_counts"])
for idx_lab in range(1, N + 2):
labels_out_list.append(labels_out == idx_lab)
centroids_list = list(stats["centroids"][:, 2])
labels = []
centroids = []
voxels = []
for idx, count in enumerate(voxel_counts_list):
if count > 10000:
labels.append(labels_out_list[idx])
centroids.append(centroids_list[idx])
voxels.append(count)
top_comps = [
(counts0, labels0, centroids0)
for counts0, labels0, centroids0 in sorted(zip(voxels, labels, centroids), reverse=True)
]
top_comps = top_comps[1:7]
# ====== Check whether the connected components are fusing vertebral bodies ======
revert_to_original = False
volumes = list(zip(*top_comps))[0]
if volumes[0] > 1.5 * volumes[1]:
revert_to_original = True
print("Reverting to original...")
labels = list(zip(*top_comps))[1]
centroids = list(zip(*top_comps))[2]
top_comps = zip(centroids, labels)
pred_centroids = [x for x, _ in sorted(top_comps)]
for label_idx in range(1, 7):
if not revert_to_original:
try:
pred_centroid = pred_centroids[6 - label_idx]
except:
# Change this part
print(
"Post processing failure, probably < 6 predicted bodies. Reverting to original labels."
)
revert_to_original = True
if revert_to_original:
pred_level = pred == label_idx
sums = np.sum(pred_level, axis=(0, 1))
indices = list(range(sums.shape[0]))
groupby_input = zip(indices, list(sums))
# sys.exit()
g = groupby(groupby_input, key=lambda x: x[1] > 0.0)
m = max([list(s) for v, s in g if v > 0], key=lambda x: np.sum(list(zip(*x))[1]))
# sys.exit()
# m = max([list(s) for v, s in g], key=lambda np.sum)
res = list(zip(*m))
indexes = list(res[0])
sums = list(res[1])
normalized_sums = sums / np.sum(sums)
pred_centroid = np.sum(indexes * normalized_sums)
print(f"Centroid for prediction {label_idx}: {pred_centroid}")
'''