Spaces:
Sleeping
Sleeping
Vemund Fredriksen
commited on
Commit
·
c7c329a
1
Parent(s):
3ff57cd
Initial pre and post implementation
Browse files- lungtumormask/dataprocessing.py +252 -0
lungtumormask/dataprocessing.py
ADDED
@@ -0,0 +1,252 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import lungmask
|
2 |
+
from lungmask import mask
|
3 |
+
from monai import transforms
|
4 |
+
from monai.transforms.intensity.array import ThresholdIntensity
|
5 |
+
from monai.transforms.spatial.array import Resize, Spacing
|
6 |
+
import torch
|
7 |
+
from tqdm import tqdm
|
8 |
+
import numpy as np
|
9 |
+
from monai.transforms import (Compose, LoadImaged, ToNumpyd, ThresholdIntensityd, AddChanneld, NormalizeIntensityd, SpatialCropd, DivisiblePadd, Spacingd, SqueezeDimd)
|
10 |
+
|
11 |
+
def mask_lung(scan_path, batch_size=20):
|
12 |
+
model = lungmask.mask.get_model('unet', 'R231')
|
13 |
+
device = torch.device('cuda')
|
14 |
+
model.to(device)
|
15 |
+
|
16 |
+
scan_dict = {
|
17 |
+
'image' : scan_path
|
18 |
+
}
|
19 |
+
|
20 |
+
transformer = Compose(
|
21 |
+
[
|
22 |
+
LoadImaged(keys=['image']),
|
23 |
+
ToNumpyd(keys=['image']),
|
24 |
+
]
|
25 |
+
)
|
26 |
+
|
27 |
+
scan_read = transformer(scan_dict)
|
28 |
+
inimg_raw = scan_read['image'].swapaxes(0,2)
|
29 |
+
|
30 |
+
tvolslices, xnew_box = lungmask.utils.preprocess(inimg_raw, resolution=[256, 256])
|
31 |
+
tvolslices[tvolslices > 600] = 600
|
32 |
+
tvolslices = np.divide((tvolslices + 1024), 1624)
|
33 |
+
|
34 |
+
torch_ds_val = lungmask.utils.LungLabelsDS_inf(tvolslices)
|
35 |
+
dataloader_val = torch.utils.data.DataLoader(torch_ds_val, batch_size=batch_size, shuffle=False, num_workers=1,
|
36 |
+
pin_memory=False)
|
37 |
+
|
38 |
+
timage_res = np.empty((np.append(0, tvolslices[0].shape)), dtype=np.uint8)
|
39 |
+
|
40 |
+
with torch.no_grad():
|
41 |
+
for X in tqdm(dataloader_val):
|
42 |
+
X = X.float().to(device)
|
43 |
+
prediction = model(X)
|
44 |
+
pls = torch.max(prediction, 1)[1].detach().cpu().numpy().astype(np.uint8)
|
45 |
+
timage_res = np.vstack((timage_res, pls))
|
46 |
+
|
47 |
+
outmask = lungmask.utils.postrocessing(timage_res)
|
48 |
+
|
49 |
+
|
50 |
+
outmask = np.asarray(
|
51 |
+
[lungmask.utils.reshape_mask(outmask[i], xnew_box[i], inimg_raw.shape[1:]) for i in range(outmask.shape[0])],
|
52 |
+
dtype=np.uint8)
|
53 |
+
|
54 |
+
outmask = np.swapaxes(outmask, 0, 2)
|
55 |
+
#outmask = np.flip(outmask, 0)
|
56 |
+
|
57 |
+
|
58 |
+
return outmask.astype(np.uint8), scan_read['image_meta_dict']['affine']
|
59 |
+
|
60 |
+
def calculate_extremes(image, annotation_value):
|
61 |
+
|
62 |
+
holder = np.copy(image)
|
63 |
+
|
64 |
+
x_min = float('inf')
|
65 |
+
x_max = 0
|
66 |
+
y_min = float('inf')
|
67 |
+
y_max = 0
|
68 |
+
z_min = -1
|
69 |
+
z_max = 0
|
70 |
+
|
71 |
+
holder[holder != annotation_value] = 0
|
72 |
+
|
73 |
+
holder = np.swapaxes(holder, 0, 2)
|
74 |
+
for i, layer in enumerate(holder):
|
75 |
+
if(np.amax(layer) < 1):
|
76 |
+
continue
|
77 |
+
if(z_min == -1):
|
78 |
+
z_min = i
|
79 |
+
z_max = i
|
80 |
+
|
81 |
+
y = np.any(layer, axis = 1)
|
82 |
+
x = np.any(layer, axis = 0)
|
83 |
+
y_minl, y_maxl = np.argmax(y) + 1, layer.shape[0] - np.argmax(np.flipud(y))
|
84 |
+
x_minl, x_maxl = np.argmax(x) + 1, layer.shape[1] - np.argmax(np.flipud(x))
|
85 |
+
|
86 |
+
if(y_minl < y_min):
|
87 |
+
y_min = y_minl
|
88 |
+
if(x_minl < x_min):
|
89 |
+
x_min = x_minl
|
90 |
+
if(y_maxl > y_max):
|
91 |
+
y_max = y_maxl
|
92 |
+
if(x_maxl > x_max):
|
93 |
+
x_max = x_maxl
|
94 |
+
|
95 |
+
return ((x_min, x_max), (y_min, y_max), (z_min, z_max))
|
96 |
+
|
97 |
+
def process_lung_scan(scan_dict, extremes):
|
98 |
+
|
99 |
+
load_transformer = Compose(
|
100 |
+
[
|
101 |
+
LoadImaged(keys=["image"]),
|
102 |
+
ThresholdIntensityd(keys=['image'], above = False, threshold = 1000, cval = 1000),
|
103 |
+
ThresholdIntensityd(keys=['image'], above = True, threshold = -1024, cval = -1024),
|
104 |
+
AddChanneld(keys=["image"]),
|
105 |
+
NormalizeIntensityd(keys=["image"]),
|
106 |
+
SpatialCropd(keys=["image"], roi_start=(extremes[0][0], extremes[1][0], extremes[2][0]), roi_end=(extremes[0][1], extremes[1][1], extremes[2][1])),
|
107 |
+
Spacingd(keys=["image"], pixdim=(1, 1, 1.5)),
|
108 |
+
]
|
109 |
+
)
|
110 |
+
|
111 |
+
processed_1 = load_transformer(scan_dict)
|
112 |
+
|
113 |
+
transformer_1 = Compose(
|
114 |
+
[
|
115 |
+
DivisiblePadd(keys=["image"], k=16, mode='constant'),
|
116 |
+
SqueezeDimd(keys=["image"], dim = 0),
|
117 |
+
ToNumpyd(keys=["image"]),
|
118 |
+
]
|
119 |
+
)
|
120 |
+
|
121 |
+
processed_2 = transformer_1(processed_1)
|
122 |
+
|
123 |
+
affine = processed_2['image_meta_dict']['affine']
|
124 |
+
|
125 |
+
normalized_image = processed_2['image']
|
126 |
+
|
127 |
+
return normalized_image, affine
|
128 |
+
|
129 |
+
def preprocess(image_path):
|
130 |
+
|
131 |
+
preprocess_dump = {}
|
132 |
+
|
133 |
+
scan_dict = {
|
134 |
+
'image' : image_path
|
135 |
+
}
|
136 |
+
|
137 |
+
im = LoadImaged(keys=['image'])(scan_dict)
|
138 |
+
preprocess_dump['org_shape'] = im['image'].shape
|
139 |
+
preprocess_dump['pixdim'] = im['image_meta_dict']['pixdim'][1:4]
|
140 |
+
preprocess_dump['org_affine'] = im['image_meta_dict']['affine']
|
141 |
+
|
142 |
+
masked_lungs = mask_lung(image_path, 5)
|
143 |
+
right_lung_extreme = calculate_extremes(masked_lungs[0], 1)
|
144 |
+
preprocess_dump['right_extremes'] = right_lung_extreme
|
145 |
+
right_lung_processed = process_lung_scan(scan_dict, right_lung_extreme)
|
146 |
+
|
147 |
+
left_lung_extreme = calculate_extremes(masked_lungs[0], 2)
|
148 |
+
preprocess_dump['left_extremes'] = left_lung_extreme
|
149 |
+
left_lung_processed = process_lung_scan(scan_dict, left_lung_extreme)
|
150 |
+
|
151 |
+
|
152 |
+
preprocess_dump['affine'] = left_lung_processed[1]
|
153 |
+
|
154 |
+
preprocess_dump['right_lung'] = right_lung_processed[0]
|
155 |
+
preprocess_dump['left_lung'] = left_lung_processed[0]
|
156 |
+
|
157 |
+
return preprocess_dump
|
158 |
+
|
159 |
+
def find_pad_edge(original):
|
160 |
+
a_min = -1
|
161 |
+
a_max = original.shape[0]
|
162 |
+
|
163 |
+
for i in range(len(original)):
|
164 |
+
a_min = i
|
165 |
+
if(np.any(original[i])):
|
166 |
+
break
|
167 |
+
|
168 |
+
for i in range(len(original) - 1, 0, -1):
|
169 |
+
a_max = i
|
170 |
+
if(np.any(original[i])):
|
171 |
+
break
|
172 |
+
|
173 |
+
original = original.swapaxes(0,1)
|
174 |
+
|
175 |
+
b_min = -1
|
176 |
+
b_max = original.shape[0]
|
177 |
+
|
178 |
+
for i in range(len(original)):
|
179 |
+
b_min = i
|
180 |
+
if(np.any(original[i])):
|
181 |
+
break
|
182 |
+
|
183 |
+
for i in range(len(original) - 1, 0, -1):
|
184 |
+
b_max = i
|
185 |
+
if(np.any(original[i])):
|
186 |
+
break
|
187 |
+
|
188 |
+
original = original.swapaxes(0,1)
|
189 |
+
original = original.swapaxes(0,2)
|
190 |
+
|
191 |
+
c_min = -1
|
192 |
+
c_max = original.shape[0]
|
193 |
+
|
194 |
+
for i in range(len(original)):
|
195 |
+
c_min = i
|
196 |
+
if(np.any(original[i])):
|
197 |
+
break
|
198 |
+
|
199 |
+
for i in range(len(original) - 1, 0, -1):
|
200 |
+
c_max = i
|
201 |
+
if(np.any(original[i])):
|
202 |
+
break
|
203 |
+
|
204 |
+
return a_min, a_max + 1, b_min, b_max + 1, c_min, c_max + 1
|
205 |
+
|
206 |
+
|
207 |
+
def remove_pad(mask, original):
|
208 |
+
a_min, a_max, b_min, b_max, c_min, c_max = find_pad_edge(original)
|
209 |
+
return mask[a_min:a_max, b_min:b_max, c_min: c_max]
|
210 |
+
|
211 |
+
def voxel_space(image, target):
|
212 |
+
image = Resize((target[0][1]-target[0][0], target[1][1]-target[1][0], target[2][1]-target[2][0]), mode='trilinear')(np.expand_dims(image, 0))[0]
|
213 |
+
image = ThresholdIntensity(above = False, threshold = 0.5, cval = 1)(image)
|
214 |
+
image = ThresholdIntensity(above = True, threshold = 0.5, cval = 0)(image)
|
215 |
+
|
216 |
+
return image
|
217 |
+
|
218 |
+
def stitch(org_shape, cropped, roi):
|
219 |
+
holder = np.zeros(org_shape)
|
220 |
+
|
221 |
+
holder[roi[0][0]:roi[0][1], roi[1][0]:roi[1][1], roi[2][0]:roi[2][1]] = cropped
|
222 |
+
|
223 |
+
return holder
|
224 |
+
|
225 |
+
def post_process(left_mask, right_mask, preprocess_dump):
|
226 |
+
left = remove_pad(left_mask, preprocess_dump['left_lung'])
|
227 |
+
right = remove_pad(right_mask, preprocess_dump['right_lung'])
|
228 |
+
|
229 |
+
left = voxel_space(left, preprocess_dump['left_extremes'])
|
230 |
+
right = voxel_space(right, preprocess_dump['right_extremes'])
|
231 |
+
|
232 |
+
left = stitch(preprocess_dump['org_shape'], left, preprocess_dump['left_extremes'])
|
233 |
+
right = stitch(preprocess_dump['org_shape'], right, preprocess_dump['right_extremes'])
|
234 |
+
|
235 |
+
stitched = np.logical_or(left, right).astype(int)
|
236 |
+
|
237 |
+
return stitched
|
238 |
+
|
239 |
+
|
240 |
+
if __name__ == "__main__":
|
241 |
+
path = "D:\\Datasets\MSD\\Images\\lung_003.nii.gz"
|
242 |
+
preprocess_dump = preprocess(path)
|
243 |
+
|
244 |
+
unpad = post_process(preprocess_dump['left_lung'], preprocess_dump['right_lung'], preprocess_dump)
|
245 |
+
|
246 |
+
import nibabel
|
247 |
+
|
248 |
+
nimage = nibabel.Nifti1Image(unpad, preprocess_dump['org_affine'])
|
249 |
+
nibabel.save(nimage, "D:\\Datasets\\stitched.nii.gz")
|
250 |
+
|
251 |
+
|
252 |
+
|