def fileopen():
filename = filedialog.askopenfilename(
filetypes=[('mhd files',
'*.mhd'), ('dicom files',
'*.dcm'), ('numpy files',
'*.npy'), ('all', '*')])
#filename = "E:/tianchi_project/TIANCHI_examples/train_1/LKDS-00001.mhd"
if filename == '':
return
print(repr(filename))
prename, extname = os.path.splitext(filename)
if extname == '.mhd':
full_image_info = sitk.ReadImage(filename)
full_scan = sitk.GetArrayFromImage(full_image_info)
old_spacing = np.array(full_image_info.GetSpacing())[::-1]
volimg, new_spacing = mt.resample(full_scan, old_spacing)
elif extname == '.dcm':
pathname = os.path.split(filename)[0]
full_scan, full_image_info, patientid = mt.read_dicom_scan(
pathname)
cvm.view_CT(full_scan)
old_spacing = np.array(full_image_info.GetSpacing())[::-1]
volimg, new_spacing = mt.resample(full_scan, old_spacing)
elif extname == '.npy':
volimg = np.load(filename)
else:
print('unknown data type')
return
label = tk.Label(tool_view, image=cvm.view_CT(volimg))
label.pack()
tool_view.quit()
def annotations_crop(
self,
randsample=True,
candsample=False,
overbound=False,
augment=False
): #the term 'augment' is invalid when 'overbound' is True
if os.access(self.output_path, os.F_OK):
shutil.rmtree(self.output_path)
os.makedirs(self.output_path)
os.mkdir(self.nodules_npy_path) #训练用正样本路径
os.mkdir(self.all_annotations_mhd_path) #检查用正样本路径
os.mkdir(self.nonnodule_npy_path) #训练用负样本路径
os.mkdir(self.no_annotation_mhd_path) #检查用负样本路径
if not os.access(self.vision_path, os.F_OK):
os.makedirs(self.vision_path)
for patient in enumerate(tqdm(self.ls_all_patients)):
patient = patient[1]
#patient = './LUNA16/subset9\\1.3.6.1.4.1.14519.5.2.1.6279.6001.114914167428485563471327801935.mhd'
print(patient)
# 检查这个病人有没有大于3mm的结节标注
if patient not in self.df_annotations.file.values:
print('Patient ' + patient + 'Not exist!')
continue
patient_uid = mt.get_serie_uid(patient)
patient_nodules = self.df_annotations[self.df_annotations.file ==
patient]
full_image_info = sitk.ReadImage(patient)
full_scan = sitk.GetArrayFromImage(full_image_info)
origin = np.array(
full_image_info.GetOrigin())[::-1] #---获取“体素空间”中结节中心的坐标
old_spacing = np.array(full_image_info.GetSpacing()
)[::-1] #---该CT在“世界空间”中各个方向上相邻单位的体素的间距
image, new_spacing = mt.resample(full_scan, old_spacing) #---重采样
print('resample done')
v_centers = []
center_coords = []
for index, nodule in patient_nodules.iterrows():
nodule_diameter = nodule.diameter_mm
nodule_center = np.array(
[nodule.coordZ, nodule.coordY,
nodule.coordX]) #---获取“世界空间”中结节中心的坐标
v_center = np.rint(
(nodule_center - origin) / new_spacing) #映射到“体素空间”中的坐标
v_center = np.array(v_center, dtype=int)
v_centers.append([index, nodule_diameter, v_center])
center_coords.append(v_center)
#volume_regioned = cvm.view_coordinations(image, center_coords, window_size=int(math.ceil(1.5*nodule_diameter)), reverse=False, slicewise=False, show=False)
#np.save(self.vision_path+"/"+patient_uid+"_annotated.npy", volume_regioned)
#---这一系列的if语句是根据“判断一个结节的癌性与否需要结合该结节周边位置的阴影和位置信息”而来，故每个结节都获取了比该结节尺寸略大的3D体素
#get annotations nodule
window_half = int(BOX_SIZE / 2)
if overbound:
num_translations = 1
for index, nodule_diameter, v_center in v_centers:
zyx_1 = v_center - BOX_SIZE # 注意是: Z, Y, X
zyx_2 = v_center + BOX_SIZE
if mt.coord_overflow(zyx_1,
image.shape) or mt.coord_overflow(
zyx_2, image.shape):
continue
nodule_box = np.zeros(
[2 * BOX_SIZE, 2 * BOX_SIZE, 2 * BOX_SIZE],
np.int16) # ---nodule_box_size = 45
img_crop = image[zyx_1[0]:zyx_2[0], zyx_1[1]:zyx_2[1],
zyx_1[2]:zyx_2[2]] # ---截取立方体
img_crop[img_crop <
-1024] = -1024 # ---设置窗宽，小于-1024的体素值设置为-1024
try:
nodule_box = img_crop[
0:2 * BOX_SIZE, 0:2 * BOX_SIZE,
0:2 * BOX_SIZE] # ---将截取的立方体置于nodule_box
except:
print("annotation error")
continue
#nodule_box[nodule_box == 0] = -1024 # ---将填充的0设置为-1000，可能有极少数的体素由0=>-1000，不过可以忽略不计
self.save_annotations_nodule(
nodule_box, patient_uid + "_" + str(index) + "_ob")
else:
if not augment:
scales = [1.0]
translations = np.array([0, 0, 0])
else:
scales = [0.8, 1.0, 1.25]
#translations = np.array([[0,0,0],[0,0,1],[0,0,-1],[0,1,0],[0,-1,0],[1,0,0],[-1,0,0]], dtype=float)
translations = np.array(
[[0, 0, 0], [0, 0, 1], [0, 0, -1], [0, 1, 0],
[0, math.sqrt(0.5), math.sqrt(0.5)],
[0, math.sqrt(0.5), -math.sqrt(0.5)], [0, -1, 0],
[0, -math.sqrt(0.5),
math.sqrt(0.5)],
[0, -math.sqrt(0.5), -math.sqrt(0.5)], [1, 0, 0],
[math.sqrt(0.5), 0, math.sqrt(0.5)],
[math.sqrt(0.5), 0, -math.sqrt(0.5)],
[math.sqrt(0.5), math.sqrt(0.5), 0],
[
math.sqrt(0.3333),
math.sqrt(0.3333),
math.sqrt(0.3333)
],
[
math.sqrt(0.3333),
math.sqrt(0.3333), -math.sqrt(0.3333)
], [math.sqrt(0.5), -math.sqrt(0.5), 0],
[
math.sqrt(0.3333), -math.sqrt(0.3333),
math.sqrt(0.3333)
],
[
math.sqrt(0.3333), -math.sqrt(0.3333),
-math.sqrt(0.3333)
], [-1, 0, 0], [-math.sqrt(0.5), 0,
math.sqrt(0.5)],
[-math.sqrt(0.5), 0, -math.sqrt(0.5)],
[-math.sqrt(0.5), math.sqrt(0.5), 0],
[
-math.sqrt(0.3333),
math.sqrt(0.3333),
math.sqrt(0.3333)
],
[
-math.sqrt(0.3333),
math.sqrt(0.3333), -math.sqrt(0.3333)
], [-math.sqrt(0.5), -math.sqrt(0.5), 0],
[
-math.sqrt(0.3333), -math.sqrt(0.3333),
math.sqrt(0.3333)
],
[
-math.sqrt(0.3333), -math.sqrt(0.3333),
-math.sqrt(0.3333)
]])
num_translations = 3
for index, nodule_diameter, v_center in v_centers:
for s in range(len(scales)):
rt = np.zeros(num_translations, dtype=int)
rt[1:num_translations] = np.random.choice(
range(1, len(translations)), num_translations - 1,
False)
rt = np.sort(rt)
for t in range(rt.size):
scale = scales[s]
box_size = int(np.ceil(BOX_SIZE * scale))
window_size = int(box_size / 2)
translation = np.array(nodule_diameter / 2 *
translations[rt[t]] /
new_spacing,
dtype=int)
tnz = translation.nonzero()
if tnz[0].size == 0 and t != 0:
continue
zyx_1 = v_center + translation - window_size # 注意是: Z, Y, X
zyx_2 = v_center + translation + box_size - window_size
if mt.coord_overflow(
zyx_1, image.shape) or mt.coord_overflow(
zyx_2, image.shape):
continue
nodule_box = np.zeros(
[BOX_SIZE, BOX_SIZE, BOX_SIZE],
np.int16) # ---nodule_box_size = 45
img_crop = image[zyx_1[0]:zyx_2[0],
zyx_1[1]:zyx_2[1],
zyx_1[2]:zyx_2[2]] # ---截取立方体
img_crop[
img_crop <
-1024] = -1024 # ---设置窗宽，小于-1024的体素值设置为-1024
if not augment or scale == 1.0:
img_crop_rescaled = img_crop
else:
img_crop_rescaled, rescaled_spacing = mt.resample(
img_crop, new_spacing, new_spacing * scale)
try:
padding_shape = (
img_crop_rescaled.shape - np.array(
[BOX_SIZE, BOX_SIZE, BOX_SIZE])) / 2
nodule_box = img_crop_rescaled[
padding_shape[0]:padding_shape[0] +
BOX_SIZE,
padding_shape[1]:padding_shape[1] +
BOX_SIZE,
padding_shape[2]:padding_shape[2] +
BOX_SIZE] # ---将截取的立方体置于nodule_box
except:
# f = open("log.txt", 'a')
# traceback.print_exc(file=f)
# f.flush()
# f.close()
print("annotation error")
continue
#nodule_box[nodule_box == 0] = -1024 # ---将填充的0设置为-1000，可能有极少数的体素由0=>-1000，不过可以忽略不计
self.save_annotations_nodule(
nodule_box, patient_uid + "_" + str(index) +
"_" + str(s * rt.size + t))
print("annotation sampling done")
#get candidate annotation nodule
candidate_coords = []
if candsample:
segimage, segmask, flag = cd.segment_lung_mask(image)
if segimage is not None:
#nodule_matrix, index = cd.candidate_detection(segimage,flag)
#cluster_labels = lc.seed_mask_cluster(nodule_matrix, cluster_size=1000)
cluster_labels = lc.seed_volume_cluster(
image, segmask, eliminate_lower_size=-1)
segresult = lc.segment_color_vision(image, cluster_labels)
cvm.view_CT(segresult)
#lc.cluster_size_vision(cluster_labels)
exit()
candidate_coords, _ = lc.cluster_centers(cluster_labels)
#candidate_coords = lc.cluster_center_filter(image, candidate_coords)
#the coordination order is [z,y,x]
print("candidate number:%d" % (len(candidate_coords)))
#volume_regioned = cv.view_coordinations(image, candidate_coords, window_size=10, reverse=False, slicewise=True, show=False)
#mt.write_mhd_file(self.vision_path+"/"+patient_uid+"_candidate.mhd", volume_regioned, volume_regioned.shape[::-1])
for cc in range(len(candidate_coords)):
candidate_center = candidate_coords[cc]
invalid_loc = False
if mt.coord_overflow(candidate_center - window_half,
image.shape) or mt.coord_overflow(
candidate_center + BOX_SIZE -
window_half, image.shape):
invalid_loc = True
continue
for index_search, nodule_diameter_search, v_center_search in v_centers:
rpos = v_center_search - candidate_center
if abs(rpos[0]) < window_half and abs(
rpos[1]
) < window_half and abs(
rpos[2]
) < window_half: #the negative sample is located in the positive location
invalid_loc = True
break
if not invalid_loc:
zyx_1 = candidate_center - window_half
zyx_2 = candidate_center + BOX_SIZE - window_half
nodule_box = np.zeros(
[BOX_SIZE, BOX_SIZE, BOX_SIZE],
np.int16) #---nodule_box_size = 45
img_crop = image[zyx_1[0]:zyx_2[0], zyx_1[1]:zyx_2[1],
zyx_1[2]:zyx_2[2]] #---截取立方体
img_crop[img_crop <
-1024] = -1024 #---设置窗宽，小于-1000的体素值设置为-1000
if img_crop.shape[0] != BOX_SIZE | img_crop.shape[
1] != BOX_SIZE | img_crop.shape[2] != BOX_SIZE:
print("error in resmapleing shape")
try:
nodule_box[
0:BOX_SIZE, 0:BOX_SIZE, 0:
BOX_SIZE] = img_crop # ---将截取的立方体置于nodule_box
except:
print("random error")
continue
#nodule_box[nodule_box == 0] = -1024#---将填充的0设置为-1000，可能有极少数的体素由0=>-1000，不过可以忽略不计
self.save_nonnodule(nodule_box,
patient_uid + "_cc_" + str(cc))
print("candidate sampling done")
#get random annotation nodule
if randsample:
if overbound:
augnum = 100
elif augment:
augnum = len(scales) * num_translations
else:
augnum = 1
if augnum * len(v_centers) > len(candidate_coords):
randnum = augnum * len(v_centers) - len(candidate_coords)
else:
randnum = len(candidate_coords)
for rc in range(
randnum
): #the random samples is one-to-one number of nodules
#index, nodule_diameter, v_center = v_centers[rc]
rand_center = np.array([0, 0, 0]) # 注意是: Z, Y, X
invalid_loc = True
candidate_overlap = True
while invalid_loc:
invalid_loc = False
candidate_overlap = False
for axis in range(rand_center.size):
rand_center[axis] = np.random.randint(
0, image.shape[axis])
if mt.coord_overflow(rand_center - window_half,
image.shape) or mt.coord_overflow(
rand_center + BOX_SIZE -
window_half, image.shape):
invalid_loc = True
continue
if 'segmask' in dir() and not (
segmask is None
) and not segmask[rand_center[0], rand_center[1],
rand_center[2]]:
invalid_loc = True
continue
for index_search, nodule_diameter_search, v_center_search in v_centers:
rpos = v_center_search - rand_center
if abs(rpos[0]) < window_half and abs(
rpos[1]
) < window_half and abs(
rpos[2]
) < window_half: #the negative sample is located in the positive location
invalid_loc = True
break
for candidate_coord in candidate_coords:
rpos = candidate_coord - rand_center
if abs(rpos[0]) < window_half and abs(
rpos[1]
) < window_half and abs(
rpos[2]
) < window_half: #the negative sample is located in the pre-extracted candidate locations
candidate_overlap = True
break
if candidate_overlap:
continue
zyx_1 = rand_center - window_half
zyx_2 = rand_center + BOX_SIZE - window_half
nodule_box = np.zeros([BOX_SIZE, BOX_SIZE, BOX_SIZE],
np.int16) #---nodule_box_size = 45
img_crop = image[zyx_1[0]:zyx_2[0], zyx_1[1]:zyx_2[1],
zyx_1[2]:zyx_2[2]] #---截取立方体
img_crop[
img_crop < -1024] = -1024 #---设置窗宽，小于-1000的体素值设置为-1000
if img_crop.shape[0] != BOX_SIZE | img_crop.shape[
1] != BOX_SIZE | img_crop.shape[2] != BOX_SIZE:
print("error in resmapleing shape")
try:
nodule_box[
0:BOX_SIZE, 0:BOX_SIZE,
0:BOX_SIZE] = img_crop # ---将截取的立方体置于nodule_box
except:
# f = open("log.txt", 'a')
# traceback.print_exc(file=f)
# f.flush()
# f.close()
print("candidate error")
continue
#nodule_box[nodule_box == 0] = -1024#---将填充的0设置为-1000，可能有极少数的体素由0=>-1000，不过可以忽略不计
self.save_nonnodule(nodule_box,
patient_uid + "_rc_" + str(rc))
print("random sampling done")
print('Done for this patient!\n\n')
print('Done for all!')
import numpy as np from toolbox import CTViewer_Multiax as cvm filename = "./detection_vision/train/LKDS-00005_detresult2.npy" volimg = np.load(filename) cvm.view_CT(volimg)