-
Notifications
You must be signed in to change notification settings - Fork 2
/
Copy pathroi_extraction.py
184 lines (168 loc) · 9.97 KB
/
roi_extraction.py
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
from scipy.ndimage import zoom
import imageio as imio
from matplotlib import path
import os
import pydicom as dicom
import numpy as np
target_size = 32
root_dir = './Varian2/'
def get_transform_matrix(sorted_images):
N = len(sorted_images)
if N<2:
return None
T1 = np.array(sorted_images[0].ImagePositionPatient).astype(np.float32)
TN = np.array(sorted_images[len(sorted_images)-1].ImagePositionPatient).astype(np.float32)
cosines = sorted_images[0].ImageOrientationPatient
delta_r = sorted_images[0].PixelSpacing[0]
delta_c = sorted_images[0].PixelSpacing[1]
F = np.array([[cosines[3],cosines[0]],[cosines[4],cosines[1]],[cosines[5],cosines[2]]])
transform = np.zeros([4,4])
transform[0:3,0]= F[:,0]*delta_r
transform[0:3,1] = F[:,1]*delta_c
transform[0:3,2] = ((T1-TN)/(1-N)).T
transform[0:3,3] = T1.T
transform[3,3]=1
return transform
def load_series(location):
# loads the dicom series and returns the sorted list of dicom image files, the rtstruct file and the average slice spacing
files = [dicom.read_file(os.path.join(location,f)) for f in os.listdir(location) if f.endswith('dcm')]
ct_files = [f for f in files if f.SOPClassUID =='1.2.840.10008.5.1.4.1.1.2']
rt_struct_file = [f for f in files if f.SOPClassUID == '1.2.840.10008.5.1.4.1.1.481.3' ]
if len(rt_struct_file)>1:
print('Found more than one structure file in the directory, will pick the first one')
elif len(rt_struct_file)==0:
print('Found no structure file in the directory')
return None, None, None
rt_struct_file = rt_struct_file[0]
files = sorted(ct_files,key=lambda x:x.SliceLocation) # basic sort (may not be the correct ordering)
# find actual ordering
orientation = np.array(files[0].ImageOrientationPatient)
vector_perp = np.cross(orientation[0:3],orientation[3:])
unit_vector_perp = vector_perp/(np.dot(vector_perp,vector_perp))**0.5
files2 = sorted(files,key=lambda x:np.dot(np.array(x.ImagePositionPatient),unit_vector_perp)) # this should be the correct ordering
# find slice spacing
slice_spacing1 = np.dot((np.array(files2[1].ImagePositionPatient) - np.array(files2[0].ImagePositionPatient)),unit_vector_perp)
avg_spacing = np.dot((np.array(files2[-1].ImagePositionPatient) - np.array(files2[0].ImagePositionPatient)),unit_vector_perp)/(len(files2)-1)
if abs(slice_spacing1-avg_spacing)>avg_spacing/(len(files2)-1):
print('maybe missing slices')
return files2,rt_struct_file,avg_spacing
def load_pixel_vol(vol,axis):
# loads the pixel array of the dicom volume and concatenates the slices along specified axis
return np.stack([image.pixel_array for image in vol],axis=axis)
def transform_coords(t_matrix,coords):
# utility to transfrom coords given a transform matrix
concat_coord = np.concatenate((coords,np.ones([coords.shape[0],1])),axis =1)
new_coords = np.dot(t_matrix,concat_coord.T)
return new_coords[0:3,:]
def get_roi_info(mask_info):
# parses an rtstruct file and returns a dictionary with the contour information
# mask_info = dicom.read_file(mask_info)
contour_dict = {}
gtv_numbers = []
for s in mask_info.StructureSetROISequence:
if 'GTV' in s.ROIName.upper():
gtv_numbers.append(s.ROINumber)
for i in range(len(mask_info.ROIContourSequence)):
if mask_info.ROIContourSequence[i].ReferencedROINumber not in gtv_numbers:
continue
contour_dict['ROIContourSequence'+str(i)] = {}
for j in range(len(mask_info.ROIContourSequence[i].ContourSequence)):
contour_data = mask_info.ROIContourSequence[i].ContourSequence[j].ContourData
coords = [[contour_data[3*i],contour_data[3*i+1],contour_data[3*i+2]] for i in range(int(len(contour_data)/3))]
sop_uid = mask_info.ROIContourSequence[i].ContourSequence[j].ContourImageSequence[0].ReferencedSOPInstanceUID
contour_dict['ROIContourSequence'+str(i)][sop_uid] = coords
return contour_dict
def process_contours(contour_dict,ref_vol):
# creates masks from each contour
voxel_2_world = get_transform_matrix(ref_vol)
x_len = ref_vol[0].pixel_array.shape[0]
y_len = ref_vol[0].pixel_array.shape[1]
z_len = len(ref_vol)
slice_grid = np.array(np.where(np.zeros([x_len,y_len])==0)).T
contour_points = []
masks = []
for roi_contour_seq in contour_dict:
new_vol = np.zeros([ref_vol[0].pixel_array.shape[0],ref_vol[0].pixel_array.shape[1],len(ref_vol)])
for sop_id in contour_dict[roi_contour_seq]:
coords = np.array(contour_dict[roi_contour_seq][sop_id])
roi_t2_pixel_coords = transform_coords(np.linalg.inv(voxel_2_world),coords)
slice_to_process = np.unique(np.round(roi_t2_pixel_coords[2:]).astype(np.int32))
if slice_to_process.shape[0]>1:
print('Slices to process:', slice_to_process)
return None
contour_path = path.Path(roi_t2_pixel_coords[0:2,:].T,closed=True)
contour_points.append(roi_t2_pixel_coords)
new_slice = np.zeros([x_len,y_len,1])
points_included = slice_grid[np.where(contour_path.contains_points(slice_grid))]
new_slice[points_included[:,0],points_included[:,1],np.zeros_like(points_included[:,1])]=1
new_vol[:,:,slice_to_process]=new_slice
masks.append(new_vol)
return masks,contour_points
def extract_roi(vol,cube_centers,pixel_cube_dims):
# extracts the roi with zero-padding if necessary
xlen, ylen, zlen = np.ceil(0.5*pixel_cube_dims).astype(np.int16)
padding = tuple([(xlen,xlen),(ylen,ylen),(zlen,zlen)])
padded_vol = np.pad(vol,pad_width=padding,mode='constant')
new_centers = np.round(np.array([xlen,ylen,zlen])+cube_centers).astype(np.int16)
return padded_vol[int(new_centers[0])-xlen:int(new_centers[0])+xlen, new_centers[1]-ylen:new_centers[1]+ylen, new_centers[2]-zlen:new_centers[2]+zlen]
def resize_tumor(vol,mask,target_size,spacings):
# resizes the roi to the fixed target_size
non_zero_locations = np.where(mask!=0)
x_min, x_max = non_zero_locations[0].min(), non_zero_locations[0].max()
y_min, y_max = non_zero_locations[1].min(), non_zero_locations[1].max()
z_min, z_max = non_zero_locations[2].min(), non_zero_locations[2].max()
cube_centers = 0.5*(np.array([x_max,y_max,z_max])+np.array([x_min,y_min,z_min]))
dims_world = np.multiply((np.array([x_max,y_max,z_max])-np.array([x_min,y_min,z_min])),spacings)
max_dimension = dims_world.max()
real_cube_dims = np.array([max_dimension]*3)
pixel_cube_dims = np.round(np.divide(real_cube_dims,spacings))
cropped_vol = extract_roi(vol,cube_centers,pixel_cube_dims)
cropped_mask = extract_roi(mask,cube_centers,pixel_cube_dims)
resized_vol = zoom(cropped_vol,zoom=target_size/np.array(cropped_vol.shape),prefilter=False)
resized_mask = zoom(cropped_mask,zoom = target_size/np.array(cropped_mask.shape),prefilter=False)
return np.stack([resized_vol,resized_mask],axis=3)
def resize_tumor2(vol, mask, target_size, spacings,intermediate_common_size = np.array([128,128,49])):
# this is the method used to resize by Mu and Ed: for each tumor, find its bounding box, for each slice, zoom the bounding box to
# 128 x 128 (do not zoom along z-axis), along the z-axis, extract only upto the bounding box edge, zero-pad to 49 slices
# finally, resize the 128 x 128 x 49 volume to 32 x 32 x 32
non_zero_locations = np.where(mask!=0)
x_min, x_max = non_zero_locations[0].min(), non_zero_locations[0].max()
y_min, y_max = non_zero_locations[1].min(), non_zero_locations[1].max()
z_min, z_max = non_zero_locations[2].min(), non_zero_locations[2].max()
pixel_dims = np.array([x_max,y_max,z_max])-np.array([x_min,y_min,z_min])
int_volume = vol[x_min:x_max,y_min:y_max,z_min:z_max]
int_target_size = np.array([intermediate_common_size[0],intermediate_common_size[1],pixel_dims[2]])
int_volume = zoom(int_volume,zoom = np.divide(int_target_size,pixel_dims),prefilter=False)
padding = tuple([(0,0),(0,0),(0,intermediate_common_size[2]-min(intermediate_common_size[2],pixel_dims[2]))])
padded_vol = np.pad(int_volume, pad_width=padding,mode='constant')
cropped_vol = padded_vol[:,:,0:intermediate_common_size[2]]
resized_vol = zoom(cropped_vol,zoom = target_size/intermediate_common_size,prefilter=False)
int_mask = mask[x_min:x_max,y_min:y_max,z_min:z_max]
int_mask = zoom(int_mask,zoom = np.divide(int_target_size,pixel_dims),mode='constant',prefilter=False)
padded_mask = np.pad(int_mask, pad_width=padding,mode='constant')
cropped_mask = padded_mask[:,:,0:intermediate_common_size[2]]
resized_mask = zoom(cropped_mask,zoom = target_size/intermediate_common_size,mode='constant',prefilter=False)
resized_mask[np.where(resized_mask>0)]=1
return np.stack([resized_vol,resized_mask],axis=3)
def normalize(vol):
return (vol - vol.min())/(vol.max()-vol.min())
def process_directory(vol_path,save_gifs=True,save_dir='./'):
vol,rt_struct_file,slice_spacing = load_series(vol_path)
if vol is None:
return []
roi_info = get_roi_info(rt_struct_file)
masks,_ = process_contours(roi_info,vol)
pixel_vol = load_pixel_vol(vol,axis=2)
spacings = np.array([vol[0].PixelSpacing[0],vol[0].PixelSpacing[1],slice_spacing])
count = 0
input_tumors = []
for mask in masks:
count+=1
input_tumor = resize_tumor2(pixel_vol,mask,target_size,spacings)
input_tumors.append(input_tumor)
if save_gifs:
normal_tumor = normalize(input_tumor[:,:,:,0])
imio.mimwrite(os.path.join(save_dir,str(count)+'.gif'),[np.concatenate([normal_tumor[:,:,i],np.ones([32,10]),input_tumor[:,:,i,1]],axis=1) for i in range(target_size)])
return input_tumors
if __name__ == '__main__':
process_directory(root_dir)